SOXEFFECT_NG(7)

🎀 SOXEFFECT_NG(7) — the effects glow-up 🎀

ok so this is soxeffect_ng(7), the manual page for the effects supported by sox_ng and libsox_ng, from the Sound eXchange_ng gang 💖 Bestie, this one's a long one, so buckle up.

✨ NAME ✨

soxeffect_ng — effects supported by sox_ng and libsox_ng 💅

🌸 SYNOPSIS 🌸

So like, besides converting, playing and recording audio files, SoX can also invoke a whole bunch of audio effects. And the way you stack them? You just specify them one after the other at the end of the SoX command line, and that forms what everyone calls an effects chain 🦋.

ngl, applying multiple effects in real time (i.e. when you're actually playing audio) may require a high performance computer. it's giving ✨demanding✨

Some of the SoX effects are primarily intended to be applied to a single instrument or voice. To make that easier, the remix effect and the global SoX option -M can be used to isolate then recombine tracks from a multitrack recording 🎚️

---

💜 MULTIPLE EFFECTS CHAINS 💜

A single effects chain is made up of one or more effects. Audio from the input runs through the chain until either the end of the input file is reached or an effect terminates the chain 🏁

SoX supports running multiple effects chains over the input audio. When one chain says "ok I'm done processing audio", the audio data gets sent through the next effects chain. This continues until either no more effects chains exist or the input has reached the end of the file 🌀

You separate effects chains by placing a : (colon) after an effect — any following effects are part of a new effects chain. slay 💅

important bestie ⚠️ put the effect that stops the chain as the first effect in the chain, because any samples that are buffered by effects to the left of the terminating effect will be discarded 💀. The amount of samples discarded is related to the --buffer option and it should be kept small, relative to the sample rate, if the terminating effect cannot be first. Further info on stopping effects lives in the Stopping SoX section.

🧚 the pseudo-effects

There are a few pseudo-effects that help when you're using multiple effects chains:

• newfile — starts writing to a new output file before moving to the next effects chain
• restart — moves back to the first effects chain

Pseudo-effects must be specified as the first effect in a chain and as the only effect in a chain (i.e. they must have a : before and after them). no exceptions, periodt 💅

🌈 example

Here's an example of multiple effects chains. It splits the input file into multiple files, each of 30 seconds in length, and each output filename will have a unique number in its name, as documented in the Output Files section:

sox_ng in.au out.au trim 0 30 : newfile : restart

---

⭐ COMMON NOTATION AND PARAMETERS ⭐

In the descriptions that follow:

• [square brackets] denote parameters that are optional
• {braces} denote those that are both optional and repeatable
• <angle brackets> denote those that are repeatable but not optional
• the pipe character | separates options from which to choose one of several alternatives

Where applicable, default values for optional parameters are shown (in parentheses) 🫶

The following parameters are used with, and have the same meaning for, several effects:

• frequency — a frequency in Hz or, if followed by k, in kHz or, if preceded by %, in semitones relative to A (440Hz); alternatively, scientific note names (e.g. E2) may be used 🎵
• gain — a power gain in dB. Zero gives no gain, less than zero gives an attenuation and greater than zero amplifies 📈
• duration — see Time Specifications below ⏱️
• position — a position within the audio stream; the syntax is [=|-|+]timespec, where timespec is a time specification (see below). The optional first character indicates whether the timespec is to be interpreted relative to the start (=) or end (-) of the audio or relative to the previous position (+) if the effect accepts multiple positional arguments. The audio length must be known for end-relative locations to work, though some effects do accept -0 for the end of the audio even if the length is unknown. Which of =, - and + is the default depends on the effect and is shown in its syntax as, e.g., position(+) 🧭

Examples: =2:00 is two minutes into the audio stream, -100s is one hundred samples before the end of the audio, +0:12+10s is twelve seconds and ten samples after the previous position and -0.5+1s is one sample less than half a second before the end of the audio 🕰️

• width[h|k|o|q] — used to specify the bandwidth of a filter. A number of different methods to specify the width are available (though not all for every effect). One of the characters shown may be appended to select the desired method as follows 🌸
  • h — Hz
  • k — kHz
  • b — Hz — Old non-frequency-warped response
  • o — octaves
  • q — Q-factor — See [2]
  • s — slope

For each effect that uses this parameter, the default method (if no character is appended) is the one that is listed first in the first line of the effect's description 👑

---

🦋 TIME SPECIFICATIONS 🦋

A timespec can be given in one the following two forms:

🕐 form one

[[hours:]minutes:]seconds[.frac][t]

For example, a time specification of 1:30.5 corresponds to one minute, thirty and ½ seconds. The component values do not have to be normalized; e.g. 1:23:45, 83:45, 79:0285, 1:0:1425, 1::1425 and 5025 are all equivalent 🤯

🔢 form two

sampless

Specifies the number of samples directly, as in 8000s. For large sample counts, e notation is supported: 1.7e6s is the same as 1700000s ✨

➕➖ chaining

Time specifications can also be chained with + or - into a new time specification where the right part is added to or subtracted from the total so far. For example, 3:00-200s means two hundred samples less than three minutes 🧮

heads up 🙈 If a time specification is a plain whole number with no t or s suffix, whether it is taken as a number of seconds or a number of samples depends on the effect in question. At present, it always means seconds except for the duration parameters of the silence effect. iykyk

---

👑 SUPPORTED EFFECTS 👑

To see whether SoX has support for an optional effect, enter sox_ng -h and look for its name in the EFFECTS list; a categorized list of the effects can be found in the accompanying README file 📖

🌟 allpass [-1|2] frequency [width[h|k|o|q]]

Apply a two-pole all-pass filter with central frequency frequency and filter width width. An all-pass filter changes the audio's frequency to phase relationship without changing its frequency to amplitude relationship. The filter is described in detail in [1] 🎀

-1 or -2 use an experimental 1-pole or 2-pole filter, in which case width does not apply.

This effect supports the --plot global option ✨

🌟 band [-n] frequency [width[h|k|o|q]]

Apply a band-pass filter. The frequency response drops logarithmically around the center frequency. The width parameter gives the slope of the drop: the frequencies at frequency + width and frequency - width will have half their original amplitudes. Its default value is half of the center frequency 💅

band defaults to a mode oriented to pitched audio, i.e. voice, singing or instrumental music. The -n (for noise) option uses the alternate mode for unpitched audio (e.g. percussion), though -n introduces a power gain of about 11dB in the filter, so beware of output clipping 😭. band introduces noise in the shape of the filter, peaking at the center frequency and settling around it.

This effect supports the --plot global option ✨

See sinc for a band-pass filter with steeper shoulders 👀

🌟 bandpass|bandreject [-c] frequency width[h|k|o|q|b]

Apply a two-pole Butterworth band-pass or band-reject filter with central frequency frequency, and (3dB-point) bandwidth width. The -c option applies only to bandpass and selects a constant skirt gain (peak gain = Q) instead of the default, a constant 0dB peak gain. The filters roll off at 6dB per octave (20dB per decade) and are described in detail in [1] 🔥

These effects support the --plot global option ✨

See sinc for a band-pass filter with steeper shoulders 👀

🌟 bass|treble gain [frequency [width[s|h|k|o|q]]]

Boost or cut the bass (lower) or treble (upper) frequencies of the audio using a two-pole shelving filter with a response similar to that of a standard hifi's tone controls. This is also known as shelving equalization 🎚️

gain gives the gain at 0Hz for bass or, for treble, whichever is the lower of ∼22kHz and the Nyquist frequency. Its useful range is about -20 (for a large cut) to +20 (for a large boost). Beware of Clipping when using a positive gain ⚠️

The filter can be fine-tuned using the following optional parameters:

• frequency — sets the filter's central frequency and so can be used to extend or reduce the frequency range to be boosted or cut. The default values are 100Hz for bass and 3kHz for treble 🎯
• width — determines how steep the filter's shelf transition is. In addition to the common width specification methods, slope (the default) may be used. Its useful range is about 0.3 for a gentle slope to 1 (the maximum) for a steep slope; and its default value is 0.5 📐

The filters are described in detail in [1].

These effects support the --plot global option ✨

See equalizer for a peaking equalization effect 👀

🌟 bend [-f frame-rate(25)] [-o oversampling(16)] {start-position(+),cents,end-position(+)}

Changes the pitch by specified amounts at specified times without changing the duration 🌀 Each given triple: start-position,cents,end-position specifies one bend. cents is the number of cents (100 cents = 1 semitone) by which to bend the pitch. The other values specify the points in time at which to start and end bending the pitch. During each bend, the frequency changes logarithmically, i.e. by the same number of cents per second 📉

The pitch bending algorithm uses the Discrete Fourier Transform (DFT) at a particular frame rate and oversampling rate. The -f (from 10 to 80) and -o (from 4 to 32) parameters may be used to adjust these parameters and thus control the smoothness of the changes in pitch ✨

For example, an initial tone is generated, then bent three times, yielding four different notes in total:

play_ng -n synth 2.5 sin 667 gain 1 \
  bend .35,180,.25  .15,740,.53  0,-520,.3

Okay so here, the first bend runs from 0.35 to 0.6 seconds and the second one from 0.75 to 1.28 seconds. Quick note, bestie: the clipping that is produced in this example is deliberate; to remove it, use gain -5 in place of gain 1.

See pitch.

🎛️ biquad b0 b1 b2 a0 a1 a2

Apply a biquad Infinite Impulse Response filter with the given coefficients, where b* and a* are the numerator and denominator coefficients respectively. Filter math era, lowkey iconic.

See http://en.wikipedia.org/wiki/Digital_biquad_filter (where a0 = 1).

This effect supports the --plot global option.

✂️ centercut [-a gain] [-b] [-w size]

Remove the center from a stereo file leaving the far left and right parts of the stereo file intact and the center in a third channel. Karaoke-prep main character energy, fr fr.

The -a option is gain-out for all channels, default 1.0.

The -b option moves the bass (below 200Hz) out of the center into the sides for those who want karaoke.

The -w option changes the window size from its default of 8192 sample frames to a power of two from 4 to 32768.

A stereo equivalent of the oops effect is

sox_ng in.wav centercut remix 1 2

The keymap for the -a option is centercut.gain.

See also oops.

📻 channels channels

Invoke a simple algorithm to change the number of channels in the audio signal to the given number: mixing if decreasing the number of channels or duplicating if increasing the number of channels. Channel glow-up, no cap.

The channels effect is invoked automatically if SoX's -c option specifies a number of channels that is different to that of the input file(s). Alternatively, if this effect is given explicitly, SoX's -c option need not be given. For example, the following two commands are equivalent:

sox_ng input.wav -c 1 output.wav bass -b 24
sox_ng input.wav      output.wav bass -b 24 channels 1

though the second form is more flexible as it allows the effects to be ordered arbitrarily.

For example, when making a stereo file quadraphonic, the left and right channels are copied into the third and fourth and when mixing a four-channel file down to stereo, the left channel is the mix of the first and third and the right of the second and fourth.

See remix for an effect that allows channels to be mixed and selected arbitrarily.

🎤 chorus [-n|l|q] [-s|t] [gain-in [gain-out {delay [decay [speed [depth [-s|t]]]]}]]

Add a chorus effect to the audio. This can make a single voice sound like a chorus but can also be applied to instrumentation. Obsessed with a good chorus, literally.

Chorus resembles an echo effect with a short delay but, while echo's delay is constant, chorus' delay varies by a sinusoidal or triangular modulation.

See [3] for further discussion of the chorus effect.

The -l flag makes chorus do linear interpolation between samples when the offset into the delay line is not a whole number, which is about 15% slower but makes it considerably less noisy and -q asks for quadratic interpolation which is about 40% slower but makes it even less noisy. -n explicitly asks for no interpolation, the default, fast and fuzzy.

-s or -t before the stages change the default wave type for all of them.

All parameters are all optional and, if missing, assume the following values:

• gain-in — range -1-1, default 0.5; proportion of input delivered clean to the adder.
• gain-out — range -1-1, default 1; final volume adjustment.
• delay — range 0-1000, default 40-60; fixed delay in milliseconds.
• decay — range -1-1, default 0.5; volume of delayed output.
• speed — range 0-192k, default 0.25; modulation frequency.
• depth — range 0-1000, default 2; extra delay in milliseconds.
• wave — range -s|-t, default -s; sinusoidal/triangular modulation.

There are keymaps for gain_in and gain_out.

Each delay ranges from the fixed delay to delay + depth.

Gain-out is then applied to the sum of the input scaled by gain-in and the outputs from the delays scaled by their decays.

For internal reasons regarding the speed of chorus, there is a limit of 256 chorus stages.

A typical delay is around 40ms to 60ms; the modulation speed is best near 0.25Hz and the modulation depth around 2ms. For example, a single delay:

play_ng guitar1.wav chorus 0.7 0.9 55 0.4 0.25 2 -t

Two delays of the original samples:

play_ng guitar1.wav chorus 0.6 0.9 50 0.4 0.25 2 -t \
   60 0.32 0.4 1.3 -s

A fuller-sounding chorus (with three additional delays):

play_ng guitar1.wav chorus 0.5 0.9 50 0.4 0.25 2 -t \
   60 0.32 0.4 2.3 -t 40 0.3 0.3 1.3 -s

flanger can do everything that chorus does except multiple stages but works in floating point internally instead of integers so is slower without a floating point processor:

chorus -l gain-in gain-out delay decay speed depth -wave

is equivalent to

flanger delay depth 0 100×decay÷gain-in speed wave 0 \
   vol gain-out÷(gain-in+decay)

For a flow diagram of how chorus works, say sox_ng --help-effect chorus.

🎚️ compand attack1,decay1{,attack,decay} [soft-knee-dB:]in-dB1[,out-dB1]{,in-dB,out-dB} [gain [initial-volume-dB [delay]]]

Compand (compress or expand) the dynamic range of the audio. Dynamics gymnastics, but make it cute.

The attack and decay parameters (in seconds) determine the time over which the instantaneous level of the input signal is averaged to determine its volume; attacks refer to increases in volume and decays refer to decreases. For most situations, the attack time (its response to the music getting louder) should be shorter than the decay time because the human ear is more sensitive to sudden loud music than sudden soft music. When more than one pair of attack/decay parameters is specified, each input channel is companded separately and the number of pairs must agree with the number of input channels. Typical values are 0.3,0.8 seconds.

The second parameter is a list of points on the compander's transfer function specified in dB relative to the maximum possible signal amplitude. The input values must be in a strictly increasing order but the transfer function does not have to be monotonically rising. If omitted, the value of out-dB1 defaults to the same value as in-dB1; levels below in-dB1 are not companded but may have gain applied to them. The point 0,0 is assumed but may be overridden by 0,out-dBn. If the list is preceded by a soft-knee-dB value, then the points at where adjacent line segments on the transfer function meet are rounded by the amount given. Typical values for the transfer function are 6:-70,-60,-20.

The third (optional) parameter is an additional gain in dB to be applied at all points on the transfer function and allows easy adjustment of the overall gain.

The fourth (optional) parameter is an initial level to be assumed for each channel when companding starts. This lets you supply a nominal level initially so that, for example, a very large gain is not applied to initial signal levels before the companding action has begun to operate: it is quite probable that in such an event, the output would be severely clipped while the compander gain adjusts itself. A typical value (for audio which is initially quiet) is -90 dB.

The fifth (optional) parameter is a delay in seconds. The input signal is analyzed immediately to control the compander, but it is delayed before being fed to the volume adjuster. Specifying a delay approximately equal to the attack/decay times allows the compander to operate in a predictive rather than a reactive mode. A typical value is 0.2 seconds.

---

The following example might be used to make a piece of music with both quiet and loud passages suitable for listening to in a noisy environment such as a moving vehicle:

sox_ng asz.wav asz-car.wav compand 0.3,1 6:-70,-60,-20 -5 -90 0.2

The transfer function (6:-70,...) says that very soft sounds (below -70dB) remain unchanged. This stops the compander from boosting the volume on silent passages such as between movements. However, sounds in the range -60dB to 0dB (maximum volume) are boosted so that the 60dB dynamic range of the original music is compressed 3-to-1 into a 20dB range, which is wide enough to enjoy the music but narrow enough to get around the road noise. The 6: selects 6dB soft-knee companding. The -5 dB output gain is needed to avoid clipping (the number is inexact and was derived by experimentation). The -90 dB for the initial volume will work fine for a clip that starts with near silence and the delay of 0.2 seconds makes the compander react more quickly to sudden volume changes.

In the next example, compand is used as a noise-gate for when the noise is at a lower level than the signal:

play_ng in.au compand .1,.2 -inf,-50.1,-inf,-50,-50 0 -90 .1

Here is another noise-gate, this time for when the noise is at a higher level than the signal (making it, in some ways, similar to a squelch effect):

play_ng in.au compand .1,.1 -45.1,-45,-inf,0,-inf 45 -90 .1

This effect supports the --plot global option (for the transfer function).

For a flow diagram of how compand works, say sox_ng --help-effect compand.

See mcompand for a multiple-band companding effect.

✨ contrast [amount(75)]

Comparable with compression, this effect modifies an audio signal to make it sound louder. amount controls the amount of the enhancement and is a number in the range 0-100. Note that amount = 0 still gives a significant contrast enhancement. Louder without all the drama? Slay.

There is a keymap for amount.

See the compand and mcompand effects.

dcshift

dcshift shift [limiter-gain]

Slap a DC shift on the audio, bestie. This is lowkey useful when you need to remove a known DC offset — maybe your recording chain has a hardware moment 💀 — because a DC offset eats your headroom and therefore your volume. The stat or stats effect can tell you if a signal has a DC offset in the first place.

The dcshift value is a floating point number in the range of ±2, and it says how much to shift the audio, which itself is in the range of ±1. You can also pass an optional limiter-gain; it should be way less than 1 (think 0.05 or 0.02) and it only kicks in on peaks to prevent clipping ✨.

Another route to removing a DC offset — with a short delay, yes — is the highpass filter effect at, say, 10Hz, like this:

sox_ng -n dc.wav synth 5 sin %0 50
sox_ng dc.wav fixed.wav highpass 10

---

deemph

deemph

Apply Compact Disc (IEC 60908) de-emphasis with a treble attenuation shelving filter 🎀.

Pre-emphasis was applied in the mastering of some CDs issued in the early 1980s — many classical music albums, plus now sought-after issues of albums by The Beatles, Pink Floyd and others. Pre-emphasis should be removed at playback time by a de-emphasis filter in the playback device. But not all modern CD players have this filter and very few PC CD drives have it; playing pre-emphasized audio without the correct de-emphasis filter makes audio sound harsh and is far from what its creators intended 😭.

With the deemph effect you can apply the necessary de-emphasis to audio extracted from a pre-emphasized CD, then either burn the de-emphasized audio to a new CD — which will then play correctly on any CD player — or just play the correctly de-emphasized audio files on the PC. For example:

sox_ng track1.wav track1-deemph.wav deemph

and then burn track1-deemph.wav to CD, or:

play_ng track1-deemph.wav

or simply:

play_ng track1.wav deemph

The de-emphasis filter is implemented as a biquad and requires the input audio sample rate to be either 44.1kHz or 48kHz. Its maximum deviation from the ideal response is only 0.06dB (up to 20kHz) 👑.

This effect supports the --plot global option.

---

delay

delay {position(=)}

Delay zero or more audio channels so they start at the given position 🌸.

For example, delay 1.5 +1 3000s delays the first channel by 1.5 seconds, the second channel by 2.5 seconds (one second more than the previous channel), the third channel by 3000 samples, and leaves other channels undelayed. The following (one long) command plays a chime sound:

play_ng -n synth -j 3 sin %3 sin %-2 sin %-5 sin %-9 \
  sin %-14 sin %-21 fade h .01 2 1.5 delay \
  1.3 1 .76 .54 .27 remix - fade h 0 2.7 2.5 norm -1

and this one gives you an arpeggiated guitar chord:

play_ng -n synth pl G2 pl B2 pl D3 pl G3 pl D4 pl G4 \
  delay 0 .05 .1 .15 .2 .25 remix - fade 0 4 .1 norm -1

With no parameters it does nothing. To delay all channels by the same amount, use the pad effect 🫶.

---

dither

dither [-S|-s|-f filter] [-a] [-p precision]

Apply dithering to the audio. Dithering deliberately adds a small amount of noise to the signal to mask audible quantization effects that can occur if the output sample size is less than 24 bits. With no options, this effect adds TPDF white noise ✨.

• -S — selects a slightly “sloped” TPDF, biased towards higher frequencies. It can be used at any sampling rate, but below ≈22kHz, plain TPDF is probably better, and above ≈37kHz, noise-shaping (if available) is probably better 💅.
• -s — enables noise-shaping with the shibata filter (the same as -f shibata).
• -f filter — lets you select a particular noise-shaping filter from this list: lipshitz, f-weighted, modified-e-weighted, improved-e-weighted, gesemann, shibata, low-shibata, high-shibata, shibata-(A|B)(0|1|2|3|4|5|6) and shibata-A-saturated. The latter shibata- ones use the new shaper coefficients from Naoki Shibata's SSRC package, described at https://shibatch.org/ssrc 🔥.

The filter types are distinguished by the following properties: audibility of noise, level of (inaudible, but in some circumstances problematic) shaped high frequency noise, and processing speed — and they are available for the following sample rates:

• lipshitz — 44100
• e- and f-weighted — 48000
• gesemann — 44100, 48000
• shibata — 8000, 11025, 16000, 22050 and 32000, 37800, 44100, 48000
• low-shibata — 44100, 48000
• high-shibata — 44100
• shibaba-A0 and A1 — 8000, 11025, 22050, 44100, 48000, 88200, 96000, 192000
• shibata-A2 — 44100, 48000, 88200, 96000, 192000
• shibata-A3 to A6 — 44100, 48000
• shibata-B0 to B6 — 44100, 48000
• shibata-A-saturated — 8000, 11025, 22050
• -a — enables a mode where dithering (and noise-shaping if applicable) are automatically enabled only when needed. The most likely use for this is when applying fade in or out to an already dithered file, so that the redithering applies only to the faded portions.
• -p precision — overrides the target precision in bits and can be from 1 to 24.

⚠️ But auto dithering is not foolproof, bestie, so the fades should be checked carefully for any noise modulation; if this occurs, then either redither the whole file or use trim and fade and concatenate the results 😭.

If the SoX global -R option is not given, the pseudo-random number generator used to generate the white noise is reseeded, i.e. the generated noise will be different on every invocation 🌈.

If the target precision is 1-bit, the sdm effect is applied automatically with default settings. Invoke it manually to control its options.

See the above section on Dithering.

---

dolbyb

dolbyb [-e|d] [-u upsamp] [-h] [-t gain(1.0)] [-a prec(-5.0)]
[-f {1|2|3|4}]

dolbyb is a Dolby B decoder/encoder based on dolbybcsoftwaredecode which simulates the operation of a Dolby B en/decoder's electronic circuit 🎧.

By default, dolbyb applies Dolby B decoding to its input signal; with -e it does Dolby B encoding. -d is also accepted but only for symmetry, as it is the default mode of operation.

• -e — do Dolby B encoding instead of decoding.
• -d — accepted for symmetry only; this is already the default decoding mode.
• -u upsamp — sets the upsampling ratio to use in the sliding filter. Digital filtering only works well if the sample rate is well above the cutoff frequency of the filter. For Dolby B's sliding filter, that frequency can be as high as 34kHz and this does not work well if the sample rate is only 44.1Khz. To get around this, it upsamples the audio to a higher rate when it passes through this filter. By default, -u0, the upsampling rate is set automatically so that the upper sample rate is at least 200Khz; upsampling can be switched off with -u1 ✨.
• -h — if given, upsampling is used throughout the effect from when the audio enters to when it leaves, not just in the sliding filter. As dolbyb's up/downsampling algorithm is simple (repeating and averaging samples) you may obtain higher quality results by upsampling with rate before dolbyb -u1 and downsampling it afterwards 💅.
• -t gain(1.0) — “threshold” adjusts the gain when the audio is fed to the Dolby gain control circuits. When a tape deck is encoding or decoding a magnetic tape, it knows the signal level at the tape heads, but with audio files the maximum signal level may not accurately represent the tape's maximum flux density (200nWb/m for cassette tapes), giving erroneous results. The -t option adjusts the volume level at which the sliding filter reacts to overcome this. Its default value is 1.0, which assumes that the maximum amplitude of the signal represents the maximum recording level on tape; higher values assume that it was recorded too quietly and values below 1.0 are for when it was recorded too loud 🫶.
• -a prec(-5.0) — in decode mode, the program has to use trial and error to get the right output sample values. -a sets how accurate it needs to be before it is considered OK. A figure of 0.0 dB would mean an accuracy of about 1 sample value. The default is -5.0 dB, which is accurate to less than one sample value.
• -f {1|2|3|4} — this one is in the synopsis; the page doesn't give further explanation here, so I'm not inventing anything 💀.

To begin with, when you have little idea of what level to use, try a wide range of levels like 5, 10, 15 and 20. If the result sounds muffled, the threshold is too low and if it seems to have too much treble, the threshold is too high. Once you know the approximate level, you can try more closely-spaced levels and listen carefully to find the best level possible. Logic would suggest listening to where tracks fade out, to see if the treble increases, but this method doesn't seem to work well and the best way seems to be to see how low the level can be set before the results sound dull and muffled, then choose a level a bit higher than this; you can just about hear the difference between results that differ in threshold setting by about 2 🌸.

The keymap for gain is dolbyb.gain. For example:

play_ng -V -k D:dolbyb.gain+2 -k d:dolbyb.gain-2 in.wav

lets you adjust the Threshold Gain in +/- 2dB steps; to see what the new value is as you proceed, use -V 👑.

• -f — picks which of four filter types to use 🎛️. So the tea: the program originally simulated an analog circuit for a Dolby B noise reducer. But too much filtering in the side path was altering the phase of the side path audio, which caused problems when the side path was recombined with the main signal. Basically, signals don't add together very well if there's too much difference in the phase 😭. To fix this, there are now 4 filter modes with hopefully less of a phase change:
  • -f1 — is the original method.
  • -f2 — is a newer method that seems to work better than 1.
  • -f3 — is another rearrangement which in practice doesn't seem to be any better than 1.
  • -f4 — seems to work best, hence it is the default mode. 👑

For further detail on these parameters and advice on digitizing and processing Dolby B-encoded tapes, consult the wiki pages at https://codeberg.org/sox_ng/libdolbyb 🫶

• dop — DSD over PCM, bestie 💿. 1-bit DSD data is packed into 24-bit samples for transport over non-DSD-aware links.
• downsample [factor(2)] — downsample the signal by an integer factor: only the first of each factor samples is retained, the others are discarded. 🗑️

No decimation filter is applied. If the input is not a properly band-limited baseband signal, aliasing will occur. This may be desirable, e.g., for frequency translation.

The new lower sample rate propagates forward in the effects chain but, unless you specify the new sample rate with -r before the output filename or with a final (no-op) rate effect, it will be resampled back up to the original sample rate.

For a general resampling effect with antialiasing, see rate. See upsample.

• earwax — this effect takes a 44.1kHz stereo signal and adds audio cues that, when listened to on headphones, move the sound stage from inside your head to outside and in front of you, as if listening to loudspeakers 🎧✨. The way it un-headphones your brain is genuinely iconic.

To see how earwax works, say sox_ng --help-effect earwax.

• echo gain-in gain-out <delay decay> — add echoes to the audio. In nature, echoes are reflected sound and digital echo effects emulate this and are often used to help fill out the sound of a single instrument or vocal 🌈.

Gain-in controls how much of the input signal is delivered clean to the output, delay is the time difference in milliseconds between the original signal and its reflection, decay is the loudness of the reflected signal and gain-out is a final volume adjustment of the result.

There are keymaps for gain_in and gain_out.

There is no limit to the number of delay/decay pairs you can use and gains and decays can be negative or greater than 1 if you wish 😌.

echo extends the length of the signal by the maximum delay time.

For example, this makes it sound as if there are twice as many instruments as are actually playing:

play_ng lead.aiff echo 0.8 0.88 60 0.4

If the delay is very short, it sound like a metallic robot: music 🤖:

play_ng lead.aiff echo 0.8 0.88 6 0.4

A longer delay sounds like an open air concert in the mountains 🏔️:

play_ng lead.aiff echo 0.8 0.9 1000 0.3

One mountain more, and:

play_ng lead.aiff echo 0.8 0.9 1000 0.3 1800 0.25

For a flow diagram of how echo works, say sox_ng --help-effect echo.

• echos gain-in gain-out <delay decay> — echos stands for Echo in Sequel and adds a sequence of echoes to the audio 💫. That is, the first echo takes the input, the second the input and the first echo, the third the input and the output of the second echo and so on. A single echos has the same effect as a single echo. Each delay decay pair gives the delay in milliseconds (with a minimum of one sample) and the decay of that echo. Gain-out is a final volume multiplier applied to the sum of the input × gain-in and the delays' outputs × their respective decays, and there are keymaps for both.

There are keymaps for gain_in and gain_out. (Yes, that's the second time, she really wants you to know about the keymaps 💅)

echos extends the length of the signal by the maximum delay time.

For example:

The sample is bounced twice in symmetric echos:

play_ng lead.aiff echos 0.8 0.7 700 0.25 700 0.3

The sample is bounced twice in asymmetric echos:

play_ng lead.aiff echos 0.8 0.7 700 0.25 900 0.3

The sample sounds as if it were played in a garage:

play_ng lead.aiff echos 0.8 0.7 40 0.25 63 0.3

For a flow diagram of how echos works, say sox_ng --help-effect echos.

• equalizer frequency width[q|o|h|k] gain — apply a two-pole peaking equalization filter 🎚️. With this filter, the signal level at and around a selected frequency can be increased or decreased while, unlike band-pass and band-reject filters, the level at all other frequencies is unchanged.

frequency gives the filter's central frequency in Hz, width gives its bandwidth and gain the required gain or attenuation in dB. Beware of clipping when using a positive gain ⚠️.

In order to produce complex equalization curves, this effect can be given several times, each with a different central frequency.

The filter is described in detail in [1].

This effect supports the --plot global option.

• fade [type] fade-in-length [stop-position(=) [fade-out-length]] — apply a fade effect to the beginning, end, or both of the audio 🌅.

An optional type can be specified to select the shape of the fade curve: q for quarter of a sine wave, h for half a sine wave, t for linear (triangular) slope, l for logarithmic, p for inverted parabola and s for square-law. The default is logarithmic.

A fade-in starts from the first sample and ramps the signal level from 0 to full volume over the time given as fade-in-length. Specify 0 if no fade-in is wanted.

For a fade-out, the audio is truncated at stop-position and the signal level is ramped from full volume down to 0 over an interval of fade-out-length before the stop-position. If fade-out-length is not specified, it defaults to the same value as fade-in-length. No fade-out is performed if stop-position is not specified. If the audio length can be determined from the input file header and any previous effects, then -0 (or, for historical reasons, 0) may be specified for stop-position to indicate the usual case of a fade out that ends at the end of the input audio stream.

See the splice effect.

• fir [coefs-file|coef <coef>] — use SoX's FFT convolution engine with given Finite Impulse Response filter coefficients 🧮. If a single argument is given, it is the name of a file containing the filter coefficients (white space separated; may contain # comments). If the filename is - or if no argument is given, the coefficients are read from the standard input (stdin); otherwise, coefficients may be given on the command line. Examples:

sox_ng in.au out.au fir .0195 -.082 .234 .891 -.145 .043

sox_ng in.au out.au fir coefs.txt

with coefs.txt containing

# HP filter: freq=10000
  1.2311233052619888e-01
 -4.4777096106211783e-01
  5.1031563346705155e-01
 -6.6502926320995331e-02

This effect supports the --plot global option.

• firfit [knots-file|<freq gain>] — use SoX's FFT convolution engine to make a filter whose frequency response approximates a spline passing through a series of frequency/gain pairs 🪢. If a single argument is given, it is the name of a file containing the knots (white space separated; may contain # comments). If the given filename is - or if no argument is given, the knots are read from the standard input (stdin); otherwise, knots may be given on the command line.

Gains are in dB and the knot frequencies must be in increasing order.

Examples:

sox_ng in.au out.au firfit 20 0 10000 -3

gives a gentle low-pass filter and

sox_ng in.au out.au firfit knots.txt

with knots.txt containing

# Approximate telephone response
300  -100
400   -10
480     0
2800    0
3000  -10
3400 -100

approximates the response of a carbon microphone telephone ☎️.

This effect supports the --plot global option.

• flanger [-n|l|q] [-s|t] [delay(0) [depth(2) [regen(0) [width(71) [speed(0.5) [shape(sine)] [phase(25) [interp(linear)]]]]]]] — apply a flanging effect to the audio 🌀. See [3] for a detailed description of flanging.

The parameters give the base delay and the added swept delay in milliseconds, the percentage of regeneration (the delayed signal feedback), width the percentage of delayed signal that is mixed with the original, speed the number of sweeps per second, the shape of the swept wave (sine or triangle), the percentage of phase shift of the swept wave in multichannel flanges (0 = 100 = the same phase on each channel) and the type of digital delay line *interp*olation (none, linear or quadratic).

sine, triangle, none, linear and quadratic can be abbreviated.

The input and the delay's output are mixed and balanced so they don't clip, so a width of 100 gives 50:50 mixing; to obtain only the delayed output and none of the input, specify width as inf.

Despite containing a delay, flanger does not extend the length of the signal so, if you also want the last dregs of the delayed output and feedback, pad the signal beforehand 😬.

sine, triangle, none, linear and quadratic can be abbreviated and -s, -t, -n, -l and -q are alternative ways to set the waveshape and the interpolation type without having to specify the rest of the parameters ✨.

For a flow diagram of how flanger works, say sox_ng --help-effect flanger.

🎚 gain — the volume knob that ate 💅

gain [-e|B|b|r] [-n] [-l|h] [gain-dB(0)]

Okay bestie, this one bumps the audio signal up or down, or — in some cases — bumps just some of its channels. 👑

Heads up fr fr: using any of -e, -B, -b, -r and -n needs temporary file space to stash the audio it's about to process, so it might not vibe with streamed audio at all. 😭

With no other options, gain-dB shifts the signal power level by that many dB: positive = amplify (beware of clipping, no cap, it will clip 😬), negative = attenuate. When you do add other options, that gain-dB amplification or attenuation gets applied after whatever those options did. Slay.

• -e — equalizes the levels of the audio channels of a multichannel file, like: gain is applied to all channels other than the one with the highest peak level, so every channel hits the same peak level. But — and this is the sneaky part 👀 — without also passing -n, the audio is not normalized.
• -B (balance) — same energy as -e, except it uses the RMS level instead of the peak level. Honestly iconic for fixing stereo imbalance caused by an imperfect record turntable cartridge. 🎶 But unlike -e, -B might cause some clipping. Not me warning you twice 💀.
• -b — like -B but with clipping protection built in: if attenuation is needed while balancing to keep clipping away, it gets applied to all channels. And when you pair it with -n, -B and -b are literally synonymous, iykyk.
• -r — only makes sense alongside a prior invocation of gain with the -h option. Scroll down for the tea. 🍵
• -n — normalizes the audio to 0dB FSD. Super often used together with a negative gain-dB so the audio lands at a given level below 0dB. Like:

sox_ng in.au out.au gain -n

normalizes to 0dB, and

sox_ng in.au out.au gain -n -3

normalizes to -3dB. chef's kiss 🎀

• -l — invokes a simple limiter. For example:

sox_ng in.au out.au gain -l 6

applies 6dB of gain but never clips. ✨ Do note: limiting more than a few dBs more than occasionally across a piece of audio is not recommended, because it can cause audible distortion. 😭 If you want a glow-up limiter, see the compand effect.

• -h — applies gain to give you headroom for whatever comes next. Example time:

sox_ng in.au out.au gain -h bass +6

Here 6dB of attenuation is applied before the bass-boosting effect, which guarantees it doesn't clip. With bass it's pretty obvious how much headroom you need, but with other effects (e.g. rate, dither) it's not always that clear. 💭 The other flex of gain -h over just typing an explicit attenuation is that if the subsequent effects don't actually use the headroom, you can reclaim it with gain -r, like:

sox_ng in.au out.au gain -h bass +6 rate 44100 gain -r

That effects chain above is guaranteed to never clip and never amplify — it attenuates if necessary to prevent clipping, but only by as much as it needs to. Periodt. 👑

Output formatting (dithering and bit-depth reduction) also needs headroom, and that headroom cannot be reclaimed, e.g.

sox_ng in.au out.au gain -h bass +6 rate 44100 gain -rh dither

Here the second gain invocation reclaims as much of the headroom as it can from the preceding effects, but hangs on to as much headroom as is needed for the subsequent processing. 🫶 And the SoX global option -G can be given to automatically invoke gain -h and gain -r for you.

Also, tiny note: synth without the -n option already incorporates the functionality of gain -h. 💅

See the norm and vol effects.

---

🌈 highpass — bye-bye low end

highpass [-1|2] frequency [width[q|o|h|k]]

Applies a high-pass filter with a 3dB point at frequency. The filter can be single-pole (with -1) or double-pole (that's the default, or say -2 out loud). width only applies to double-pole filters; default is Q = 0.707, which gives you a Butterworth response. 🧈 These filters roll off at 6dB per pole per octave (20dB per pole per decade). The double-pole filters are described in detail in [1].

This effect supports the --plot global option. 📈

See sinc if you want filters with a steeper roll-off.

---

🌀 hilbert — the 90-degree plot twist

hilbert [-n taps]

Applies an odd-tap Hilbert transform filter, phase-shifting the signal by 90 degrees. Used in many matrix coding schemes and for analytic signal generation. The process is often written as a multiplication by i (or j), the imaginary unit. 🧚

An odd-tap Hilbert transform filter has a band-pass character — it attenuates the lowest and highest frequencies. Its bandwidth can be controlled by the number of filter taps, which by default is chosen for a cutoff frequency of about 75 Hz. For a cutoff frequency of about N Hz, give the -n option with the sample rate divided by N. The number of taps can be from 3 to 1,073,741,823 — but the max value requires 56GB of physical RAM to finish within minutes rather than days, and 100,000,001 requires 18GB. 💀 Like, plan accordingly bestie.

This effect supports the --plot global option. 📈

---

🔌 ladspa — plug in, plug in 🎧

ladspa [-l] [-r] module [plugin] {argument}

Applies a LADSPA [5] (Linux Audio Developer's Simple Plugin API) plugin. Despite the name, LADSPA is not Linux-specific, and there's a whole wide range of effects available as LADSPA plugins — like CMT [6] (the Computer Music Toolkit) and Steve Harris's plugin collection [7]. The first argument is the plugin module, the second is the name of the plugin (a module can contain more than one plugin), and any other arguments are for the control ports of the plugin. Missing arguments get supplied by default values if that's possible. ✨

Normally the number of input ports of the plugin must match the number of input channels, and the number of output ports determines the output channel count. But: the -r (replicate) option lets you clone a mono plugin so it can handle multichannel input. Clone wars, but make it audio. 👯

Some plugins introduce latency, which SoX may optionally compensate for. The -l (latency compensation) option automatically compensates for latency as reported by the plugin via an output control port named "latency".

If it's set, the environment variable LADSPA_PATH is used as the search path for plugins. See LADSPA_PATH in the section ENVIRONMENT.

---

🔊 loudness — gain, but make it human

loudness [gain [reference]]

Loudness control is similar to the gain effect but provides equalization for the human auditory system. 🧠 See http://en.wikipedia.org/wiki/Loudness for a detailed description of loudness. The gain is adjusted by the given gain parameter (usually negative) and the signal is equalized according to ISO 226 w.r.t. a reference level of 65dB — though you can give an alternative reference level if the original audio was equalized at some other level. If you don't give a gain value, a default gain of -10dB is used. 🎀

See the gain effect.

---

🌸 lowpass — the mirror image

lowpass [-1|2] frequency [width[q|o|h|k]]

Applies a low-pass filter. See the description of the highpass effect for details — it's the same lore, just flipped. 🪞

---

🎛 mcompand — companding, but make it multiband

mcompand "compand-args" {frequency "compand-args"}

The quoted compand-args are exactly as for the compand effect:

attack1,decay1{,attack,decay}
[soft-knee-dB:]in-dB1[,out-dB1]{,in-dB,out-dB}
[gain [initial-volume-dB [delay]]]

The multi-band compander is similar to the single-band compander, but the audio first gets divided into bands using Linkwitz-Riley crossover filters, and then a separately specifiable compander is run on each band. 🎚 See the compand effect for the definition of its parameters. Compand parameters go between double quotes, and the crossover frequency for that band is given by crossover-freq; you can repeat these to create multiple bands. Iconic.

The following examples approximate Dolby A compression and decompression, as used for tape noise reduction in professional recording studios:

# Dolby A compressor
sox_ng in.au dolbyA.au mcompand \
   ".1,.1 4:-56,-46,-36,-26,-26,-20,-17,-15,-9,-9" 80 \
   ".1,.1 4:-56,-46,-36,-26,-26,-20,-17,-15,-9,-9" 3k \
   ".1,.1 4:-56,-46,-36,-26,-26,-20,-17,-15,-9,-9" 9k \
   ".1,.1 4:-56,-42,-36,-23,-26,-18,-17,-14,-9,-9"

# Dolby A decompressor
sox_ng dolbyA.au out.au mcompand \
   ".1,.1 4:-46,-56,-26,-36,-20,-26,-15,-17,-9,-9" 80 \
   ".1,.1 4:-46,-56,-26,-36,-20,-26,-15,-17,-9,-9" 3k \
   ".1,.1 4:-46,-56,-26,-36,-20,-26,-15,-17,-9,-9" 9k \
   ".1,.1 4:-42,-56,-23,-36,-18,-26,-14,-17,-9,-9"

Real Dolby A probably compands each channel separately, but that is left as an exercise to interested readers. 📝 (not me doing homework 💀)

See compand for a single-band companding effect.

---

🔍 noiseprof — study the noise before you fight it

noiseprof [profile-file]

Calculates a profile of the audio for use in noise reduction. See the description of the noisered effect for the details. 🕵️‍♀️

---

🧹 noisered — yeet the hiss

noisered [profile-file [amount]]

Reduces noise in the audio signal by profiling and filtering. This effect is moderately effective at removing consistent background noise such as hiss or hum — so like, manage your expectations bestie, but it's still cute. 💅

To use it, first run SoX with the noiseprof effect on a section of audio that ideally would contain silence but in fact contains noise — those sections are typically found at the beginning or the end of a recording. noiseprof writes a noise profile to profile-file, or to stdout if there's no profile-file or if - is given. E.g.

sox_ng speech.wav -n trim 0 1.5 noiseprof speech.noise-profile

Then, to actually remove the noise, run SoX again — this time with the noisered effect. noisered reduces noise according to a noise profile generated by noiseprof, from profile-file if it's given, or from stdin if there's no profile-file or if - is given. E.g.

sox_ng speech.wav cleaned.wav noisered speech.noise-profile 0.3

How much noise gets yeeted is up to amount — a number from 0 to 1, and it defaults to 0.5. ✨ Bigger numbers = more noise gone, but also a bigger chance of snatching away parts of the audio you actually wanted to keep. 😭 So before you overwrite your original recording with the noise-reduced one, please please please experiment with different amount values to find the one that's optimal for your audio. Put on headphones 🎧 and check you're happy with the results, paying particular attention to the quieter sections of the audio — that's where it shows.

On most systems, the two stages — profiling and reduction — can be combined using a pipe, like:

sox_ng noisy.wav -n trim 0 1 noiseprof | \
   play_ng noisy.wav noisered

---

norm 💅

norm [dB-level(0)]

Normalize the audio. That's it, that's the whole vibe. norm is just an alias for gain -n, so if you want the deets, go read the gain effect. 🫶

---

oops 🎤

Out Of Phase Stereo effect. It mixes stereo to twin mono, where each mono channel contains the difference between the left and right stereo channels. This is sometimes known as the karaoke effect because it often has the effect of removing most or all of the vocals from a recording. 🎶✨ It is equivalent to remix 1,2i 1,2i.

---

overdrive 🔥

overdrive [gain(20) [color(20)]]

Non-linear distortion, no cap. The color parameter controls the amount of even harmonic content in the overdriven output. Both parameters range from 0 to 100.

There are keymaps for both gain and color.

See also centercut.

---

pad 🎀

pad { [%]length[@position(=)] }

Pad the audio with silence — at the beginning, at the end, or at any specified points throughout the audio. Iconic behaviour, honestly. length is the amount of silence to insert, and position is the position in the input audio stream at which to insert it. You can specify any number of lengths and positions, provided that a specified position is not less that the previous one. position is optional for the first and last lengths specified, and if omitted they correspond to the beginning and the end of the audio respectively.

For example, pad 1.5 1.5 adds 1.5 seconds of silence at each end of the audio, whilst pad 4000s@3:00 inserts 4000 samples of silence 3 minutes into the audio. If silence is wanted only at the end of the audio, either specify the end position or specify a zero-length pad at the start.

If a pad specification starts with a % sign, the output is padded to a multiple of length at the specified position. For example, pad 0 %10 adds silence at the end of the audio up to the next multiple of 10 seconds. 🦋

See delay for an effect that can add silence at the beginning of the audio on a channel-by-channel basis.

---

phaser 🌈

phaser [-n|l|q] [-s|t] [gain-in(.4) gain-out(.74) delay(3) regen(.4)
speed(.5) [-s|t]

Add a phasing effect to the audio. See [3] for a detailed description of phasing.

delay gives the maximum delay in milliseconds from 0 to 1000, regen the amount of feedback from the delay from −1 to +1, and speed the frequency of delay-time modulation wave in Hz.

The modulation is either sinusoidal (-s, the default), which is preferable for multiple instruments, or triangular (-t) which gives single instruments a sharper phasing effect. regen can be from −1 to +1 but should usually be less than 0.5 to avoid clipping, and gain-out is the final volume adjustment from −1 to +1.

The -l flag makes phaser do linear interpolation between samples when the offset into the delay line is not a whole number — which is about 15% slower but much less noisy 😌 — and -q does quadratic interpolation, which is about 50% slower but even less noisy. -n explicitly asks for no interpolation, the default: fast and fuzzy.

In sox_ng, -s or -t can be given at the start or the end; to be compatible with earlier versions of SoX, supply all the parameters with one of these at the end, use gain-in and gain-out from 0 to 1, delay from 0 to 5, speed from 0.1 to 2 and don't use interpolation.

There are keymaps for gain_in, gain_out and regen.

⚠️ Technically, the SoX phaser is not a phaser; it is a flanger. A flanger does comb filtering with equidistant spacing (e.g. 100Hz, 200Hz, 300Hz, 400Hz, ...), while a real phaser does comb filtering with factored spacing (e.g. 100Hz, 200Hz, 400Hz, 800Hz, ...) that sounds more harmonic. The more you know 💀

For example:

play_ng snare.flac phaser 0.8 0.74 3 0.4 0.5 -t

Gentler:

play_ng snare.flac phaser 0.9 0.85 4 0.23 1.3 -s

A popular sound:

play_ng snare.flac phaser 0.89 0.85 1 0.24 2 -t

More severe:

play_ng snare.flac phaser 0.6 0.66 3 0.6 2 -t

For a flow diagram of how phaser works, say sox_ng --help-effect phaser. 🧚

---

pitch 🎵

pitch [-q] shift [segment [search [overlap]]]

Change the audio pitch but not the tempo. Main character energy.

shift gives the pitch shift as positive or negative cents (i.e. 100ths of a semitone).

Heads up bestie 🙈: raising the pitch increases the sample rate and this can make following effects slower, in particular pitch or tempo thenselves, whose running times are proportional to the sample rate times the overlap, all squared. This can be compensated for by following pitch with a fast rate effect.

pitch and tempo share the same fundamental algorithm; see the tempo effect for a description of the other parameters.

See the bend, speed and tempo effects.

---

rate 🚀

rate [-q|l|m|g|h|e|v|u] [override-options] [frequency]

Change the audio sampling rate (i.e. resample the audio) to any given frequency — even non-integer if this is supported by the output file format — using a quality level defined as follows:

• -q — quick: bandwidth n/a, rejection ≈30@Fs/4 — typical use: playback on ancient hardware 💀
• -l — low: bandwidth 80%, rejection 100 — typical use: playback on old hardware
• -m — medium: bandwidth 95%, rejection 100 — typical use: audio playback
• -g — generic: bandwidth 95%, rejection 100 — typical use: 16-bit
• -h — high: bandwidth 95%, rejection 125 — typical use: 20-bit for 16-bit mastering
• -e — extreme: bandwidth 95%, rejection 150 — typical use: 24-bit
• -v — very high: bandwidth 95%, rejection 175 — typical use: 28-bit for 24-bit mastering
• -u — ultra: bandwidth 95%, rejection 200 — typical use: 32-bit

These can also be selected with -Q n with n from 0 to 7.

B/W (bandwidth) is the percentage of the audio frequency band that is preserved, and Rej dB is the level of noise rejection. Increasing levels of resampling quality come at the expense of increasing amounts of time to process the audio. ⏳ If no quality option is given, the quality level used is high when processing audio and low when playing it. See Playing & Recording Audio above.

The quick algorithm uses cubic interpolation; all others use band-limited interpolation. By default, all algorithms have a linear phase response; for medium and above, the phase response is configurable (see below).

The rate effect is invoked automatically if SoX's -r option specifies a rate that is different to that of the input file(s). Alternatively, if this effect is given explicitly, then SoX's -r option need not be given. For example, the following two commands are equivalent:

sox_ng input.wav -r 48k output.wav bass -b 24
sox_ng input.wav        output.wav bass -b 24 rate 48k

though the second command is more flexible as it allows rate options to be given, and allows the effects to be ordered arbitrarily. 💅

A user notes that resampling tracks and then concatenating them is more likely to create clicks at the joints than joining them first and resampling the result, due to edge effects. Noted! 📝

Override Options ⚙️

The simple quality selection described above provides settings that satisfy the needs of the vast majority of resampling tasks. Occasionally, however, it may be desirable to fine-tune the resampler's filter response; for qualities medium and above, this can be achieved using the override options in the following table:

• -M/-I/-L — Phase response=minimum/intermediate/linear
• -s — Steep filter (bandwidth=99%)
• -a — Allow aliasing/imaging above the pass band
• -b width — Any bandwidth % (74-99.7 or 85-99.7 with 0
• -p phase — Any phase response (0=minimum, 25=intermediate, 50=linear, 100=maximum)

All resamplers use filters that can sometimes create echo (a.k.a. ringing) artefacts with transient signals such as those that occur with finger snaps or other highly percussive sounds. Such artefacts are much more noticeable to the human ear if they occur before the transient (pre-echo) than if they occur after it (post-echo). Note that the frequency of any such artefacts is related to the smaller of the original and new sampling rates, but if this is at least 44.1kHz, the artefacts will lie outside the range of human hearing. 👂✨

A phase response setting may be used to control the distribution of any transient echo between pre and post: with minimum phase, there is no pre-echo but the longest post-echo; with linear phase, pre- and post-echo are in equal amounts (in signal terms, but not in audibility); the intermediate phase setting attempts to find the best compromise by selecting a small length (and level) of pre-echo and a medium-length of post-echo.

A minimum, intermediate or linear phase response is selected using the -M, -I and -L options; a custom phase response can be created with the -p option. Note that phase responses between linear and maximum (greater than 50) are rarely useful. 🫶

So a resampler's bandwidth setting is literally the whole vibe check for how much frequency content survives the conversion — measured against the original sample rate when you're upsampling, or the new sample rate when you're downsampling. The pass band is the term for everything up to that bandwidth point. Like, for a 44.1kHz sample rate with a resampling bandwidth of 95%, the pass band is everything from 0Hz (DC) up to circa 21kHz. 🎀

Bumping the bandwidth up = slower conversion, and it can also serve up extra transient echo artefacts (and yeah, the reverse is true too).

The -s steep filter option swings the resampling bandwidth from the default 95% (that's based on the 3dB point, bestie) to 99%. And -b lets you set the bandwidth to literally any value in the range 74-99.7% — but heads up, values greater than 99% are not recommended for normal use because they can cause excessive transient echo. 🙈

If the -a option is in the mix, aliasing/imaging above the pass band is allowed. So with 44.1kHz and a bandwidth of 95%, that means frequency content above 21kHz can get distorted. But since that's above the pass band (i.e. above the highest frequency you actually care about / can hear), honestly it may not be a problem. The perks of allowing aliasing/imaging? Reduced processing time and transient echo artefacts reduced by almost half. 🔥

The -d option sets the bit-accuracy in the range 15 to 33, or -R sets the bit-accuracy to obtain rejection of a specified number of dB.

Examples, because examples slay:

sox_ng input.wav -b 16 output.wav rate -s -a 44100 dither -s

That's default (high) quality resampling with overrides for a steep filter, to allow aliasing, at a 44.1kHz sample rate and noise-shaped dithering to a 16-bit WAV file. 👑

sox_ng input.wav -b 24 output.aiff rate -v -I -b 90 48k

And that one is very high quality resampling with overrides for an intermediate phase, a bandwidth of 90%, at a 48k sampling rate and storing the output to a 24-bit AIFF file. ✨

🦋 Advanced Options

• -i — forces the use of a particular interpolator coefficient from -1 to 2.
• -c — tries to limit the number of coefficients to a number of kilobytes; its argument can be from 100 up.
• -B — sets the percentage of the pass-band to preserve, from 53 to 95.
• -A — sets the percentage of the bandwidth without aliasing, from 85 to 100.
• -f — sets zero pass-band roll-off instead of 0.01dB for -Q 0-2.
• -n — disables internal small-integer optimizations and
• -t — increases the irrational ratio accuracy. 💅

---

🌈 remix

remix [-a|m] [-p] <out-spec>
       out-spec  = 0 | in-spec{,in-spec}
       in-spec   = [in-chan][-[in-chan2]][vol-spec]
       vol-spec  = p|i|v[volume]

Select and mix input audio channels into output audio channels, bestie. Each output channel gets specified in turn by a given out-spec, which is a list of contributing input channels plus volume specifications. 🎧

Okay so not to be confused with the -m global option, where multiple files are mix-combined before entering the effects chain. This effect operates on the audio channels within the SoX effects processing chain. Different thing entirely! 🙅‍♀️

An out-spec contains comma-separated input channel numbers and hyphen-delimited channel number ranges; alternatively, 0 may be given to create a silent output channel. For example:

sox_ng input.wav output.wav remix 6 7 8 0

creates an output file with four channels, where channels 1, 2, and 3 are copies of channels 6, 7, and 8 in the input file, and channel 4 is just silence. 🤫

Whereas:

sox_ng input.wav output.wav remix 1-3,7 3

creates a (somewhat bizarre, ngl) stereo output file where the left channel is a mix-down of input channels 1, 2, 3 and 7, and the right channel is a copy of input channel 3.

When you specify a range of channels, the channel numbers to the left and right of the hyphen are optional and default to 1 and to the number of input channels respectively. So:

sox_ng input.wav output.wav remix -

performs a mix-down of all input channels to mono. 💜

By default, where an output channel is mixed from multiple input channels, each input channel is scaled by a factor of ¹/n. Custom mixing volumes can be set by following a given input channel (or range of input channels) with a vol-spec (volume specification). That's one of the letters p, i, or v, followed by a volume number — and the meaning of that number depends on the letter:

• p — power adjust in dB. 0 = no change.
• i — power adjust in dB. Same as for p but it inverts the audio.
• v — voltage multiplier. 1 = no change; 0.5 ≈ 6dB attenuation; 2 ≈ 6dB gain; -1 = invert.

If an out-spec includes at least one vol-spec, then by default the ¹/n scaling is not applied to any other channels in that same out-spec (though it maybe still applies in other out-specs). The -a (automatic) option can be given to retain the automatic scaling in this case. Example time:

sox_ng input.wav output.wav remix 1,2 3,4v0.8

gives you channel level multipliers of 0.5,0.5 and 1,0.8. Whereas:

sox_ng input.wav output.wav remix -a 1,2 3,4v0.8

gives you channel level multipliers of 0.5,0.5 and 0.5,0.8. 🫶

The -m (manual) option disables all automatic volume adjustments, so:

sox_ng input.wav output.wav remix -m 1,2 3,4v0.8

gives channel level multipliers of 1,1 and 1,0.8.

The volume number is optional, and omitting it corresponds to no volume change. However! The only case where that's actually useful is in conjunction with i. For example, if input.wav is stereo, then:

sox_ng input.wav output.wav remix 1,2i

is a mono equivalent of the oops effect. And:

play file.mp3 remix 1,2 1i,2i

lets you get twice as much power out of a mono speaker connected between the left and right poles than you'd get from two connected the usual way. But mind it doesn't blow the amplifier, because it draws twice as much current. 😭 ⚡

If the -p option is given, any automatic ¹/n scaling is replaced by ¹/√n (power) scaling. That gives a louder mix, but one that may occasionally clip. 🔥

One iconic use of remix is splitting an audio file into a set of files, each containing one of the constituent channels, so you can do subsequent processing on individual audio channels. When more than a few channels are involved, a script like this one is useful:

#! /bin/sh
chans=`soxi_ng -c "$1"`
while [ $chans -ge 1 ]; do
   chans0=`printf %02i $chans`   # 2 digits hence up to 99 chans
   out=`echo "$1" | sed "s/\(.*\)\.\(.*\)/\1-$chans0.\2/"`
   sox_ng "$1" "$out" remix $chans
   chans=`expr $chans - 1`
done

If you gave it a file input.wav containing six audio channels, the script would produce six output files: input-01.wav, input-02.wav, ..., input-06.wav. ✨

See the swap effect.

---

🎀 repeat

repeat [count(1)|-]

Repeat the entire audio count times, or once if count isn't given. The special value - requests infinite repetition. It requires temporary file space to store the audio that's going to be repeated. And note that repeating once yields two copies: the original audio and the repeated audio. 🌀

---

🌸 reverb

reverb [-w] [reverberance(50%) [HF-damping(50%) [room-scale(100%)
        [stereo-depth(100%) [pre-delay(0ms) [wet-gain(0dB)]]]]]]

Add reverberation to the audio using the freeverb algorithm. A reverberation effect is sometimes desirable for concert halls that are too small, or that contain so many people that the hall's natural reverberance is diminished. Applying a small amount of stereo reverb to a dry mono signal usually makes it sound more natural. See [3] for a detailed description of reverberation. 🏛️

This effect increases the volume of the audio and keeps reverberating even after the input finishes. So to prevent clipping and keep the audible part of the final reverberation, a typical invocation might be:

play_ng dry.au gain -3 pad 0 1 reverb

The -w option can be given to select only the wet signal, which lets it be processed further, independently of the dry signal. E.g.

play_ng -m in.au "|sox_ng in.au -p reverse reverb -w reverse"

for a reverse reverb effect. 💫

---

🦋 reverse

Reverse the audio completely. Requires temporary file space to store the audio to be reversed. 🙃

---

💿 riaa

Apply RIAA vinyl playback equalization. The sampling rate must be 44.1, 48, 88.2, 96 or 192kHz. 🎶

This effect supports the --plot global option.

---

🔥 saturation

saturation [type [blend [offset [drive|color|threshold]]]]

Add saturation, which can produce effects ranging from subtle warmth to crunchy fuzz. The type parameter selects the saturation type: tanh (the default), sqrt or diode. 🍓

For all types, the blend parameter (default 1) controls the mixture of wet and dry signals in the output, with 1 being fully wet. The offset parameter (default 0) adds a DC offset to the input to produce asymmetric distortion. The offset is removed from the output, so that a zero input level produces a zero output level — but when the input is non-zero, the output waveform is likely to be asymmetric.

The tanh saturation type uses the hyperbolic tangent function to apply soft clipping. The drive parameter (default 1) controls the input gain, and thus the amount of distortion. 💅

The sqrt saturation type uses a mixture of two functions: x*sqrt(|x|) and sgn(x)*sqrt(|x|), which give different tonal qualities to the output. The color parameter (default 0.5) controls the mixture of these functions, with 0 being purely x*sqrt(|x|) and 1 being purely sgn(x)*sqrt(|x|). 🌈

The diode saturation type models the effect of using a pair of diodes to clip the signal when it exceeds a threshold (default 0.5). The blend parameter, by mixing the wet and dry signals, effectively controls the amount of attenuation that occurs above the threshold, from no attenuation when blend is 0 to complete attenuation (hard clipping) when blend is 1.

There are keymaps for blend, offset, drive, color and threshold.

See the overdrive effect for another kind of non-linear distortion. When the offset parameter is used to produce asymmetric distortion, the highpass effect can be used to rebalance the waveform's positive and negative amplitude.

---

💿 sdm

sdm [-f filter] [-t order] [-n num] [-l latency]

Apply a 1-bit sigma-delta modulator producing DSD output. The input should be previously upsampled, e.g. with the rate effect, to a high rate, 2.8224MHz for DSD64. The -f option selects the noise-shaping filter from the following list where the number indicates the order of the filter:

clans-4   sdm-4
clans-5   sdm-5
clans-6   sdm-6
clans-7   sdm-7
clans-8   sdm-8

The noise filter may be combined with a partial trellis/viterbi search by supplying the following options:

• -t order — Trellis order, max 32.
• -n num — Number of paths to consider, max 32.
• -l latency — Output latency, max 2048.

⚠️ The result of using these parameters is hard to predict and can include high noise levels or instability. Caution is advised, bestie.

---

🔇 silence

silence [-l] above-periods [duration threshold[d|%]]
        [below-periods duration threshold[d|%]]

Removes silence from the beginning, middle or end of the audio, where silence is determined by a specified threshold.

The above-periods value is used to indicate whether audio should be trimmed at the beginning of the audio. A value of zero indicates that no silence should be trimmed from the beginning, in which case duration and threshold are omitted. When a non-zero above-periods is specified, you must also specify a duration and threshold, and it trims audio until it finds non-silence. It will normally be 1 when trimming silence from the beginning of the audio, but it can be increased to higher values to trim all audio up to the Nth non-silence period. For example, if you have an audio file with two songs that each contains 2 seconds of silence before the song, you could specify an above-period of 2 to strip out both silences and the first song.

duration indicates the amount of time for which non-silence must be detected before it stops trimming the silence before it. By increasing duration, short bursts of quiet noise can be treated as silence and trimmed off. duration has the peculiarity that a bare number is interpreted as a sample count, not as a number of seconds. To specify seconds, either use the t suffix (as in 2t), a decimal point (as in 2.0) or specify minutes too (as in 0:02).

threshold indicates the maximum sample value in any channel that is considered silence. For digital audio, a value of 0 may be fine, but for audio recorded from analog you may wish to increase the value to include background noise. threshold numbers may be suffixed with d to indicate that the value is in decibels or % to indicate a percentage of the maximum possible sample value. By default, it is in percent.

To trim silence from the end of the audio, specify a below-periods count, which means to remove all audio after the last onset of silence is detected. Normally, this will be 1, but it can be increased to leave shorter periods of silence and the audio that follows them intact. For example, if you have a track with 1 second of silence in the middle and 1 second at the end, you could set below-period to 2 to leave the middle silence and what follows it and remove from the final silence on.

When below-periods is given, its duration specifies the length of silence that must exist before audio is not copied any more. By specifying a higher duration, shorter silences that are wanted can be left in the audio. For example, if you have a song with 1 second of silence in the middle and 2 seconds of silence at the end, a duration of 2 could be used to skip over the middle silence and trim the end instead of starting trimming from half way through.

Unfortunately, the length of the silence at the end has to be longer than any preceding silence for this to work, so you must know the length of the silence at the end.

A more reliable way to trim silence from the end is to use the silence effect in combination with the reverse effect. By first reversing the audio, you can use the above-periods to trim from what looks like the front of the file, then reverse it again to get back to normal.

To remove silence from the middle of a file, give a negative below-periods. This value is then treated as a positive value and is also used to indicate that the effect should restart processing as specified by the above-periods, making it suitable for removing periods of silence in the middle of the audio.

• -l — indicates that below-periods' duration of silent audio should be left intact at the beginning of each period of silence. For example, if you want to remove long pauses between words but do not want to remove the pauses completely.

The following example shows how this effect can be used to make a recording that does not contain the silence that usually occurs between pressing the record button and the start of the performance:

rec_ng parameters filename other-effects silence 1 5 2%

This example should remove the start of the recording until there's a period of non-silence longer than 0.2s and louder than 0.1%, then start searching for a silence that's longer than 1s and quieter than 3% and remove it if found, leaving the first 1s of it in place, then start copying again until a silence is found that's longer than 1s and quieter than 3%, trim that to 1s and so on.

sox_ng in.au out.au silence -l 1 0.2 0.1% -1 1.0 3%

---

🦋 sinc

sinc [-a att|-b beta] [-p phase|-M|-I|-L] [-t tbw|-n taps]
     [freqHP][-freqLP [-t tbw|-n taps]] [-r]] [-d]]

Apply a kaiser-windowed low-pass, high-pass, band-pass or band-reject filter to the signal. The freqHP and freqLP parameters give the frequencies of the 6dB points of a high-pass and low-pass filter that may be invoked individually or together. If both are given, freqHP less than freqLP creates a band-pass filter and freqHP greater than freqLP creates a band-reject filter. For example, the invocations

sinc 3k
sinc -4k
sinc 3k-4k
sinc 4k-3k

create a high-pass, low-pass, band-pass and band-reject filter respectively. Iconic, honestly.

The default stop band attenuation of 120dB can be overridden with -a; alternatively, the kaiser window's beta parameter can be given directly with -b.

The default transition bandwidth of 5% of the total band can be overridden with -t (and tbw in Hertz); alternatively, the number of filter taps can be given directly with -n and is limited to the range of 11-1,073,741,823 though the maximum requires 56GB of physical RAM if it is to complete in minutes rather than days, while 100,000,000 requires 8GB.

If both freqHP and freqLP are given, a -t or -n option given to the left of the frequencies applies to both frequencies; one of these options given to the right of the frequencies applies only to freqLP.

• -p, -M, -I, -L — control the filter's phase response; see the rate effect for details.
• -r — controls whether the filter should round the number of taps to the closest integer instead of truncating it.
• -d — specifies that, if a low-pass filter is being created and the cutoff frequency is at or above the Nyquist frequency, the sinc effect should be deleted from the effects chain instead of failing.

This effect supports the --plot global option. ✨

---

🎀 softvol

softvol [volume(1.0) [double-time(0) [headroom(0)]]]

The soft volume effect applies a simple multiplier to the audio ensuring that it does not clip. When a sample would have clipped the volume multiplier is automatically reduced to compensate.

It is a simple compander with the advantages of running fast, having no pre- or post-echo and reacting on the crests of the wave, so its volume-reduction glitches don't add audible noise. Lowkey slay.

volume sets the initial volume multiplier; the default of 1.0 means no change.

double-time says that the volume should slowly increase at a rate that makes it double every double-time seconds. A good value for usual music is 10 and the default value of 0 says that the volume should not increase automatically.

headroom is in dB and limits the loudest amplitude to less than the 32-bit maximum. This may be necessary when the final bit-depth reduction and/or dithering make it clip. A value of 0.1 is sufficient to protect down to a bit-depth of 8 with dithering.

There are keymaps for volume, double_time and headroom.

When playing sound in interactive mode, the v and V keys reduce and increase the volume if there is a softvol in the effects chain. If there are more than one, which one it adjusts is probably random. Not me guessing, that's literally the page. 💀

---

🌈 spectrogram

spectrogram [options]

Create a spectrogram of the audio. The audio is passed unmodified through the SoX processing chain. This effect is optional - type sox_ng --help and check the list of supported effects to see if it has been included. 🫶

...the spectrogram gets rendered as a Portable Network Graphic (PNG) file, and it's giving data-viz realness 💅 — time lives on the X axis, frequency on the Y axis, and the magnitude of the audio signal is the Z axis, shown via the color (or, if you want, the intensity) of the pixels in the X-Y plane. If your audio has multiple channels, they stack top to bottom starting at channel 1 — which is the left channel for stereo. iykyk 🎀

So like, if my.wav is a stereo file, then

sox_ng my.wav -n spectrogram

makes a spectrogram of the entire file in the file spectrogram.png. But ngl, most of the time you only care about a smaller portion of the audio, right? e.g. with

sox_ng my.wav -n remix 2 trim 20 30 spectrogram

the spectrogram shows info from only the second (right) channel, over thirty seconds of audio starting at twenty seconds in. Obsessed. ✨

And if you wanna analyze a small slice of the frequency domain, you can use the rate effect, e.g.

sox_ng my.wav -n rate 6k spectrogram

which lets you do a detailed analysis of frequencies up to 3kHz (that's half the sampling rate) — literally where the human auditory system is most sensitive 🫶. See also the -R option below.

Then with

sox_ng my.wav -n trim 0 10 spectrogram -x 600 -y 200 -z 100

those options control the size of the spectrogram's X, Y & Z axes — in this case the spectrogram area of the produced image will be 600 by 200 pixels, and the Z axis range will be 100 dB. But heads up bestie: the produced image includes axes, legends etc. and will be larger than the spectrogram size you specified — unless you gave the -r option. If each spectrogram is x × y and there are c channels, the image will be x + 144 by (y × c) + 78, plus c − 1 if -a was not given, and 20 pixels higher than that if you gave -t Title. A raw spectrogram will be x by y × c. ikr, math 💀

In this next example

sox_ng -n -n synth 6 tri 10k:14k spectrogram -z 100 -w kaiser

an analysis window with high dynamic range is selected so the spectrogram of a swept triangular wave looks its absolute best 🌈. For a similar vibe, append the following to the chime command in the description of the delay effect (above):

rate 2k spectrogram -X 200 -Z -10 -w kaiser

Options are also available to control the appearance (color set, brightness, contrast etc.) and the filename of the spectrogram; e.g. with

sox_ng my.wav -n spectrogram -m -l -o print.png

a spectrogram is created that's suitable for printing on a black and white printer 🖨️✨

---

🎀 Options 🎀

• -x num — changes the (maximum) width (X axis) of the spectrogram from its default value of 800 pixels to a number you pick, anywhere between 100 and a million. See -X and -d.
• -X num — X axis pixels per second. The default is auto-calculated to fit the audio to the X axis size if its duration is known or given with -d, or 100 otherwise. If you give this without a -x option when the length of the audio is known, this option determines the width of the spectrogram; otherwise, it affects the duration of the spectrogram. num can be from 1 (low time resolution) to 5000 (high time resolution) and doesn't have to be an integer. SoX may make a slight adjustment to the number you give, for processing quantization reasons — and if it does, SoX reports the actual number used (viewable when the SoX global option -V is in effect). 👀
• -y num — sets the size of the Y axis per channel in pixels; this is the number of frequency bins used in the Fourier analysis that produces the spectrogram. By default the Y axis size is chosen automatically, depending on the -Y height and the number of channels, with a minimum of 64.

The DFT size is set to 2 × (num − 1) and if SoX was compiled with FFTW, sizes of 2^a × 3^b × 5^c × 7^d × 11^e × 13^f where e + f < 2 are said to be fastest. If it wasn't compiled with FFTW, anything other than powers of two is a hundred of times slower, in which case heights of a power of two plus one will be faster. ⚡

• -Y num — sets the total height of the spectrogram(s). Default value is 550 pixels and the maximum is a million. If num is not an exact multiple of the number of channels with -r, the actual total height of the spectrogram area will be a few pixel rows less. For non-raw spectrograms, instead, the height of the graph area will be slightly less for the same reason, slightly more for the single-pixel row between adjacent channel graphs if -a wasn't given, and the overall height of the image will be greater by the time axes (28) and by the title (20) if present. 🧮
• -z num — Z axis (color) range in dB, default 120. This sets the dynamic range of the spectrogram to be -num dBFS to 0 dBFS. Num may range from 20 to 180. Decreasing the dynamic range effectively increases the contrast of the spectrogram display, and vice versa.
• -Z num — sets the upper limit of the Z axis in dBFS. A negative num effectively increases the brightness of the spectrogram display, and vice versa.
• -n — normalizes the upper limit of the Z axis so that the loudest pixels are shown using the brightest color in the palette — basically a kind of automatic -Z flag 💖.
• -q num — sets the Z axis quantization, i.e. the number of different colors (or intensities) to render Z axis values in. A small number (e.g. 4) gives a poster-like effect, making it easier to discern magnitude bands of similar level, and results in a smaller PNG file. The number you give specifies the number of colors to use in the Z axis range; two colors are reserved to represent out-of-range values. 🎨
• -w name — select a window function: Hann (the default), Hamming, Bartlett, Rectangular, Kaiser or Dolph. The spectrogram is produced using the Discrete Fourier Transform (DFT) algorithm and a significant parameter of this algorithm is the choice of window function. By default, SoX uses the Hann window, which has good all-round properties for frequency resolution and dynamic range. For better frequency resolution but lower dynamic range, select a Hamming window; for higher dynamic range but poorer frequency resolution, select a Dolph window.
• -W num — window adjustment parameter. This can be used to make small adjustments to the Kaiser and Dolph windows. A positive number (up to ten) increases its dynamic range, a negative number decreases it.
• -s — allow slack overlapping of DFT windows. This can, in some cases, increase image sharpness and give greater adherence to the -x value — but at the expense of a little spectral loss. 🙈
• -a — suppress the display of the axis lines. This is sometimes useful in helping to discern artefacts at the spectrogram edges.
• -r — raw spectrogram: suppress the display of axes and legends.
• -g — when there are many seconds, show them as MM:SS or HH:MM:SS.
• -m — creates a monochrome spectrogram (the default is color).
• -h — selects a high-color palette which is less visually pleasing than the default color palette, but it may make it easier to differentiate different levels. If this option is used in conjunction with -m, the result is a hybrid monochrome/color palette. 🦋
• -p num — permute the colors in a color or hybrid palette. The num parameter, from 1 (the default) to 6, selects the permutation.
• -l — creates a printer-friendly spectrogram with a light background (the default has a dark background).
• -A — selects an alternative, fixed color set. This is provided only for compatibility with spectrograms produced by another package. It should not normally be used as it has some problems — not least, a lack of differentiation at the bottom end which results in masking of low-level artefacts.
• -t text — set the image title, the text to display above the spectrogram. If you need it to be chorus or some other effect's name, surround it by spaces inside double quotes. 💬
• -c text — set (or clear) the image comment: the text to display below and to the left of the spectrogram.
• -o file — the name of the spectrogram output PNG file, default spectrogram.png. If - is given, the spectrogram is sent to the standard output (stdout).
• -L — plot the frequency on a logarithmic axis.
• -R L:H — specify the frequency range (from L to H). The frequencies can have an optional suffix

sox_ng mymusic.mp3 -n spectrogram -L -R 100:8k

By default, the lowest frequency is 0Hz for a linear graph or 1Hz for a logarithmic graph, and the highest is the Nyquist frequency.

• -i — interpolate vertically: where there are more output pixels than frequency bins, use a weighted average of the bins above and below the pixel's frequency; and where there are more frequency bins than pixels, average the bins that fall in this pixel row. 🧚

---

✨ Advanced Options ✨

So in order to process a smaller section of audio without affecting other effects or the output signal (unlike when the trim effect is used), the following options may be used: stay tuned bestie 💜

• -d duration — okay so this sets the X axis resolution so audio with the given duration (a time specification) fits the selected (or default) X axis width. It defaults, if the audio length is known, to the audio length minus the start time. Like, for example:

sox_ng input.mp3 output.wav -n spectrogram -d 1:00 stats

This creates a spectrogram showing the first minute of the audio, while the stats effect is applied to the entire audio signal. Not me loving that split-brain behavior 💅 See -X for an alternative way of setting the X axis resolution.

• -S position(=) — start the spectrogram at the given point in the audio stream. For example:

sox_ng input.aiff output.wav spectrogram -S 1:00

creates a spectrogram showing all but the first minute of the audio (the output file, however, receives the entire audio stream). Slay.

For the ability to perform off-line processing of spectral data, see stat -freq.

---

🌀 speed factor[c]

Adjust the audio speed (pitch and tempo together). factor is either the ratio of the new speed to the old speed (greater than 1 speeds it up, less than 1 slows it down) or, if the letter c is appended, it’s the number of cents (100ths of a semitone) by which the pitch (and tempo) should be adjusted: greater than 0 increases, less than 0 decreases. iykyk ✨

Technically, the speed effect only changes the sample rate information, leaving the samples themselves untouched. The rate effect is invoked automatically to resample to the output sample rate, using its default quality/speed. For higher quality or higher speed resampling, in addition to the speed effect, specify the rate effect with the desired quality option.

See the bend, pitch and tempo effects. the crossover episode we deserve 🫶

---

🧼 speexdsp

speexdsp [-agc [target_level(100)]] [-denoise [max_db(15)]] [-dereverb]
         [-fps frames_per_second(20)] [-spf samples_per_frame]

Use the Speex DSP library to improve perceived sound quality. bestie, it’s giving glow-up for your ears.

If no options are specified, the -agc and -denoise features are enabled.

• -agc [target_level] — enable automatic gain control and optionally specify a target volume level from 1 to 100. The synopsis default is 100.
• -denoise [max_db] — enable noise reduction and optionally specify the maximum attenuation from 1 to 100. The synopsis default is 15.
• -dereverb — enable reverb reduction.
• -fps frames_per_second — specify the number of frames per second from 1-100. The synopsis default is 20.
• -spf samples_per_frame — specify the number of samples per frame. The default is derived from the -fps setting so that frames abut but do not overlap.

---

✂️ splice

splice  [-h|t|q] {position(=)[,excess[,leeway]]}

Splice audio sections together. This effect provides two things over simple audio concatenation: a (usually short) cross-fade is applied at the join and a wave similarity comparison is made to help determine the best place at which to make the join. obsessed with the precision tbh 🎀

One of the options -h, -t, or -q may be given to select the fade envelope as half cosine wave (the default), triangular (a.k.a. linear), or quarter cosine wave (e.g. for a cross-fade of correlated audio).

• -h — correlated audio, constant gain, smooth transitions.
• -t — correlated audio, constant gain, abrupt transitions.
• -q — uncorrelated audio, constant power, smooth transitions.

To perform a splice, first use the trim effect to select the audio sections to be joined together. As when performing a tape splice, the end of the section to be spliced onto should be trimmed with a small excess (default 0.005 seconds) after the ideal joining point. The beginning of the audio section to splice on should be trimmed with the same excess before the ideal joining point plus an additional leeway (default 0.005 seconds). SoX should then be invoked with the two audio sections as input files and the splice effect given with the position at which to perform the splice - this is length of the first audio section (including the excess).

The following diagram uses the tape analogy to illustrate the splice operation. The effect simulates the diagonal cuts and joins the two pieces:

                   length1   excess
                 -----------><--->
                 _________   :   :  _________________
                          \  :   : :\     `
                           \ :   : : \     `
                            \:   : :  \     `
                             *   : :   * - - *
                              \  : :   :\     `
                               \ : :   : \     `
                 _______________\: :   :  \_____`____
                                   :   :   :     :
                                   <--->   <----->
                                   excess  leeway

where * indicates the joining points.

For example, a long song begins with two verses which start (as determined e.g. by using the play_ng command with the trim (start) effect) at times 0:30.125 and 1:03.432. The following commands cut out the first verse:

sox_ng too-long.wav part1.wav trim 0 30.130

(5 ms excess, after the first verse starts)

sox_ng too-long.wav part2.wav trim 1:03.422

(5 ms excess plus 5 ms leeway, before the second verse starts)

sox_ng part1.wav part2.wav just-right.wav splice 30.130

For another example, the SoX command

play_ng "|sox_ng -n -p synth 1 sin %1" "|sox_ng -n -p synth 1 sin %3"

generates and plays two notes, but there is a nasty click at the transition; the click can be removed by splicing instead of concatenating the audio, i.e. by appending splice 1 to the command. Clicks at the beginning and end of the audio can be removed by preceding the splice effect with fade q .01 2 .01.

Provided your arithmetic is good enough, multiple splices can be performed with a single splice invocation. For example, with a Bourne shell script acpo:

#! /bin/sh
# Audio Copy and Paste Over
# acpo infile copy-start copy-stop paste-over-start outfile
# No chained time specifications allowed for the parameters
# (i.e. such that contain +/-).
e=0.005                      # Using default excess
l=$e                         # and leeway.
sox_ng "$1" piece.wav trim $2-$e-$l =$3+$e
sox_ng "$1" part1.wav trim 0 $4+$e
sox_ng "$1" part2.wav trim $4+$3-$2-$e-$l
sox_ng part1.wav piece.wav part2.wav "$5" \
   splice $4+$e +$3-$2+$e+$l+$e

two splices are used to copy and paste audio. the way this is basically audio scrapbooking 😭

It is also possible to use this effect to perform general cross-fades, e.g. to join two songs. In this case, excess would typically be a number of seconds, the -q option would typically be given to select an equal power cross-fade and leeway should be zero (which is the default if -q is given). For example, if f1.wav and f2.wav are audio files to be cross-faded, then

sox_ng f1.wav f2.wav out.wav splice -q $(soxi_ng -D f1.wav),3

cross-fades the files where the point of equal loudness is 3 seconds before the end of f1.wav, i.e. the total length of the cross-fade is 2 × 3 = 6 seconds ($(...) is POSIX shell notation that is replaced by the output of the enclosed command).

---

📊 stat

stat [-s scale] [-rms] [-freq] [-v] [-d] [-a] [-h]

Display time and frequency domain statistical information about the audio. Audio is passed unmodified through the SoX processing chain.

The information is output to the standard error (stderr) stream and is calculated (where n is the duration of the audio in samples, c is the number of audio channels, r is the audio sample rate and xk represents the value (in the range -1 to +1) of each successive sample in the audio), as follows:

• Samples read — n×c
• Length (seconds) — n÷r
• Scaled by — see -s below.
• Maximum amplitude — max(xk). The maximum sample value in the audio; usually this will be a positive number.
• Minimum amplitude — min(xk). The minimum sample value in the audio; usually this will be a negative number.
• Midline amplitude — ½min(xk)+½max(xk)
• Mean norm — ¹/nΣ│xk│. The average of the absolute value of each sample in the audio.
• Mean amplitude — ¹/nΣxk. The average of each sample in the audio. If this figure is non-zero, then it indicates the presence of a DC offset which could be removed using the dcshift effect.
• RMS amplitude — √(¹/nΣxk²). The level of a DC signal that would have the same power as the audio’s average power.
• Maximum delta — max(│xk-xk-1│)
• Minimum delta — min(│xk-xk-1│)
• Mean delta — ¹/n-1Σ│xk-xk-1│
• RMS delta — √(¹/n-1Σ(xk-xk-1)²)
• EBUR128 Momentary — the maximum momentary loudness over 400ms
• EBUR128 Short Term — the maximum short term loudness over 3 seconds
• EBUR128 Integrated — the integrated loudness over the whole file
• EBUR128 True Peak — the maximum of the True Peak of each channel
• Rough frequency — in Hz.
• Volume Adjustment — the parameter to the vol effect which would make the audio as loud as possible without clipping. See the discussion on Clipping above for reasons why it is rarely a good idea actually to do this.

Note that the delta measurements are not applicable to multichannel audio and EBU R 128 (=ITU-R BS.1770) measurements are in Loudness Units referenced to Full Scale (LUFS). the math is mathing, bestie 🧚

• -s — scales the input data by a given factor 💅. The default value of scale is 2147483647 (the max value of a 32-bit signed integer, because internal effects always work with those). A lower value means a different sample value should be considered the full-scale amplitude.
• -rms — converts all average values to root mean square format ✨.
• -freq — outputs the input's power spectrum (a 4096-point DFT) instead of the statistics listed above. Bestie, this should only be used with a single-channel audio file, fr fr.
• -v — displays only the Volume Adjustment value. That's it. That's the whole vibe.
• -d — displays a hex dump of the 32-bit signed PCM data audio in SoX's internal buffer 👀. Mainly used to help track down endian problems that sometimes occur in cross-platform versions of SoX. Detective mode, lowkey.
• -a — outputs the average power spectrum instead of the power spectrum for each 4096-point DFT 🌈.
• -h — uses the "histogram algorithm" to calculate the integrated EBU R-128 loudness, which requires less memory but is less accurate. Trade-off era 💀.
• -j — outputs the statistics in JSON format, e.g.:

{
  "samples_read": 22699008,
  "length": 236.448,
  "scaled_by": 2.14748e+09,
  "maximum_amplitude": 0.818604,
  "minimum_amplitude": -0.532471,
  "midline_amplitude": 0.143066,
  "mean_norm": 0.0352694,
  "mean_amplitude": 0.00180676,
  "rms_amplitude": 0.056726,
  "maximum_delta": 0.367126,
  "minimum_delta": 0,
  "mean_delta": 0.0177341,
  "rms_delta": 0.0268538,
  "rough_frequency": 3616,
  "volume_adjustment": 1.22159
}

If -rms was given, "scaled_by" will be "scaled_by_rms" and if -e was given, you also get:

  "ebur128_momentary": -30.3408,
  "ebur128_short_term": -35.4501,
  "ebur128_integrated": -21.3583,

Some fields may be absent if their values are incalculable (EBUR128 figures) or would be infinite (like the RMS of silence) 🫥.

As JSON uses scientific notation, it can show the values of very small numbers that the usual output shows as zero ✨.

The most common use of stat is to measure the characteristics of a single audio file, for which the syntax is:

sox_ng file.wav -n stat

where -n means "No audio output is required." 🎀

---

🦋 stats

stats [-b bits|-x bits|-s scale] [-w time] [-j]

Display time domain statistical information about the audio channels; audio is passed unmodified through the SoX processing chain. Statistics are calculated and displayed for each audio channel and, where applicable, an overall figure is also given 💖.

For example, for a typical well-mastered stereo music file:

                Overall     Left      Right
   DC offset   0.000803 -0.000391  0.000803
   Min level  -0.750977 -0.750977 -0.653412
   Max level   0.708801  0.708801  0.653534
   Pk lev dB      -2.49     -2.49     -3.69
   RMS lev dB    -19.41    -19.13    -19.71
   RMS Pk dB     -13.82    -13.82    -14.38
   RMS Tr dB     -85.25    -85.25    -82.66
   Crest factor       -      6.79      6.32
   Flat factor     0.00      0.00      0.00
   Pk count           2         2         2
   Bit-depth      16/16     16/16     16/16
   Num samples    7.72M
   Length s     174.973
   Scale max   1.000000
   Window s       0.050

DC offset, Min level, and Max level are shown, by default, in the range ±1. If the -b (bits) option is given, these three measurements are scaled to a signed integer with the given number of bits from 2 to 32. For example, for 16 bits, the scale would be -32768 to +32767 ✨. The -x option behaves the same way as -b except that the signed integer values are displayed in hexadecimal. The -s option scales the three measurements by a given floating point number.

Pk lev dB and RMS lev dB are the standard peak and RMS levels measured in dBFS. RMS Pk dB and RMS Tr dB are peak and trough values of the RMS level measured over a short window (default: 50ms). That can be changed with the -w option in seconds from 0.01 to 10.

Crest factor is the ratio of peak to RMS level (note: not in dB 👑).

Flat factor is a measure of the flatness (i.e. consecutive samples with the same value) of the signal at its peak levels (i.e. either Min level or Max level).

Pk count is the number of occasions (not the number of samples) that the signal attained either Min level, or Max level. The primary goal of the Peak Count value is to answer the question "has this audio been clipped?", quite possibly as a result of the frowned-upon-by-some but common practice of 'brick wall limiting' in modern mastering 💀. The closer the "Peak Count" is to 1, the higher the confidence that the audio has not been clipped.

The right-hand Bit-depth figure is the standard definition of bit-depth, i.e. that all bits other than this number of the most significant bits are always zero. The left-hand figure is the number of bits at the least significant end of those most significant bits that would be sufficient to represent all sample values accurately (including the sign bit) 🧚.

In mathematical terms, the right-hand figure is the ordinal, counting from the most significant bit, of the least significant bit that is set to one in at least one sample. The left-hand figure is the ordinal, counting from the least significant repeated sign bit across all samples, of the least significant bit that is set to one in at least one sample.

Bit-depths are not intended to be properties of the signal per se but properties of its 2's-complement PCM encoding 🌸.

The primary use case of bit-depth measurement concerns manipulation of PCM audio by simple bit shifting, to answer questions such as: "Is it likely that this 24-bit PCM file was created by simply converting a 16-bit PCM file to 24-bit?" or "Can I losslessly shift all the samples in this PCM audio file m-bits left or n-bits right?"

For multichannel audio, an overall figure for each of the above measurements is given and derived from the channel figures as follows: DC offset: maximum magnitude; Max level, Pk lev dB, RMS Pk dB, Bit-depth: maximum; Min level, RMS Tr dB: minimum; RMS lev dB, Flat factor, Pk count: average; Crest factor: not applicable 💅.

Length s is the duration in seconds of the audio and, unlike stat, Num samples is equal to the sample rate multiplied by Length. Scale max is the scaling applied to the first three measurements; specifically, it is the maximum value that could apply to Max level. Window s is the length of the window used for the peak and trough RMS measurements.

The -j option outputs JSON with three fields:

• "channel_count" — an integer.
• "overall" — an object with a member for each row of the first column of the usual output, which are all numbers except for "bit_depth", which is an array of two numbers.
• "channels" — an array of objects with the per-channel values.

To know the overall and the channels' member names, have a look at the output 👀.

Like stat, the usual way to measure the characteristics of a single audio file is:

sox_ng file.wav -n stats

---

🎀 stretch

stretch [factor [window [fade [shift [fading]]]]]

Change the audio duration but not its pitch by cross-fading between short windows of samples ✨. This effect is broadly equivalent to the tempo effect with factor inverted and search set to zero so, in general, its results are comparatively poor; it is retained as it can sometimes outperform tempo for small factors.

factor determines the change in length: >1 lengthens and <1 shortens. By default, it is 1 (no change).

window is the length of the cross-fading window in milliseconds with a default of 20.

The fade option chooses the type of crossfading: linear and half-cosine give equal-gain crossfading and cannot clip; sqrt and quarter-cosine give two kinds of equal-power crossfading 💖.

The shift ratio can be from 0 to 1 and its default depends on the stretch factor: 1 when speeding up, 0.8 when slowing down.

The fading ratio, from 0 to 0.5, seems to be how much of each window is cross-faded with the adjacent ones. The default value depends on factor and shift: 1.0 − (factor × shift) if speeding up, 1.0 − shift if slowing down, with a maximum of 0.5 🌸.

The duration of stretch's output is slightly longer than the duration of the input multiplied by factor as it has to empty the delay line it uses; tempo is more precise 😭.

---

🌈 swap

Swap stereo channels 💅. If the input is not stereo, pairs of channels are swapped and a possible odd last channel is passed through. E.g., for seven channels, the output order will be 2, 1, 4, 3, 6, 5, 7.

See remix for an effect that allows arbitrary channel selection, ordering and mixing.

---

⭐ synth

synth [-j|p key] [-n] [length [offset [phase [p1 [p2 [p3]]]]]] {type
[combine [fixed[,extra[,mix]]]] [freq[:|+|/|-freq2] [offset [phase [p1
[p2 [p3]]]]]]}

synth generates fixed or swept frequency audio tones with various wave shapes and wide-band noise of various colors 🔥. Multiple synth effects can be cascaded to produce more complex waveforms and at each stage it is possible to choose whether the generated waveform is mixed with or modulated onto the output of the previous stage, and the audio for each channel in a multichannel audio file can be synthesized independently 👑

It generates audio at maximum volume (0dBFS), which means that there is a high chance of clipping so, in many cases, you will want to follow it with the gain effect to prevent this from happening. (See Clipping above.) ⚠️

okay so, even though this effect is literally for generating audio, you still have to give it an input file — bestie, yes, even then. The input file’s characteristics are what set the synthesized audio length, the number of channels, and the sampling rate. But since the input file’s audio is not normally needed, a null file (with the special input filename -n) is often given instead, and the length is specified as a parameter to synth or by some other effect that has an associated length.

By default, the tuning used with note notations is equal temperament; the -j key option selects just intonation, where key is a whole number of semitones relative to A (so for example, -9 or 3 selects the key of C) or a note in scientific notation, and -p selects Pythagorean tuning.

By default, the synth effect incorporates the functionality of gain -h (see the gain effect for details); synth’s -n option may be given to disable this behaviour.

length is the length of audio to synthesize. A value of 0 means use the input length, which is also the default. Note that, if the input is -n and the length is 0 or absent, it continues generating audio until it is stopped in some other way.

🚨 heads up, bestie: -n + length of 0 or absent = it keeps going until something else stops it. iconic but also kinda unhinged.

type is one of:

• sine — A sinusoidal wave. This is the default type and it ignores all the p parameters.
• square — A square wave. p1 sets the percentage of each cycle that is on, with a default of 50.

|_______        | +1
|       |       |
|_______|_______|  0
|       |       |
|       |_______| -1
|               |
0       p1      1

• triangle — p1 sets the percentage of each cycle that is rising, with a default of 50.

|    .    | +1
|   / \   |
|__/___\__|  0
| /     \ |
|/       \| -1
|         |
0    p1   1

• sawtooth — A sawtooth wave. With a phase of 0 it starts at -1 and rises to 1, and of 10 it starts at -0.9. The offset makes no difference.

|    /| +1
|   / |
|__/__|  0
| /   |
|/    | -1
0     1

• trapezium — The trapezoidal wave starts at -1, rises linearly to 1, stays there, falls linearly to -1, stays there and repeats. p1 sets the percentage of the cycle in which the wave is rising with a default of 10, p2 sets the percentage through each cycle at which falling begins with a default of 50, and p3 sets the percentage through each cycle at which falling ends with a default of 60.

|    ______             |+1
|   /      \            |
|__/________\___________| 0
| /          \          |
|/            \_________|-1
|                       |
0   p1    p2   p3       1

• exp — The exponential wave rises from -1 to 1 where it peaks and immediately begins an exponential fall. p1 sets the position of the maximum with a default of 50. p2 sets the minimum amplitude in multiples of 2dB down from the maximum with a default of 50 (100dB); values below 50 raise the shoulders of the wave and values above 50 lower the shoulders, increasing the pointedness of the spike.

|                           | +1
|            /\             |
|          _'  `_           | 0
|        _-      -_         |
|____---'          `---____ | f(p2)
|                           |
0             p1             1

• whitenoise — Random noise with equal power at every frequency. All noise generators ignore the frequency and phase parameters, but if a DC offset is given, the signal’s amplitude is automatically adjusted to prevent clipping, so for noise in the range 0 to 1, an offset of 0.5 would give a signal ranging from 0.0 to 1.0, and -0.9 from -1.0 to -0.8.

noise is a handy alias for whitenoise.

• tpdfnoise — Noise with a Triangular Probability Density Function.
• pinknoise — Random noise with the power at each frequency inversely proportional to the frequency.
• brownnoise — Random noise with the power at each frequency inversely proportional to the frequency squared.
• pluck — A plucked string simulation in which an array of sample values representing a taut string is set in motion with a burst of noise and decayed over time.

A plucked note’s frequency can be from 27.5 to 4220Hz and the sampling rate must be between 44100 and 48000Hz.

If a DC offset is used, the amplitude is automatically adjusted to prevent clipping.

p1 affects the sustain with a default of 40 (2dB per second); higher values give a slower decay and lower values a faster one.

p2 and p3 are tone controls for the initial excitation, with default values of 20 and 90, and a special case when p3 is exactly 100. If the phase is non-zero, it uses a different kind of random numbers.

If the offset, phase and p parameters are given before the first type, they set the default values for all the following stages.

combine is one of:

• create — Puts each stage’s output in a new output channel and is the default.
• mix — Mixes the generated audio 50:50 with the input signal.
• amod — Amplitude-modulates (multiplies) the input signal by the synthesized one considered as a value from 0 (for the most negative value) to 1 (for the most positive value).
• fmod — Multiplies the input signal with the synthesized one (ring modulation).
• vdelay — Mixes the input signal with a delayed version of it using the synthesized signal to modulate the depth of the delay. The following three-part option fixed[,extra[,mix]] specifies the fixed and additional parts of the delay in milliseconds and what percentage of the output consists of the delayed signal from 0 for all input signal to 100 for all delayed signal, with a default of 50 (half and half).

The synthesized signal’s value from -1 to +1 varies the delay from fixed seconds to fixed + (0 to extra) seconds.

It interpolates linearly between the input samples and can be used to make precision phaser, flanger and chorus-like effects, vibrato and frequency modulation (FM) synthesis (actually phase modulation, as used in the Yamaha DX7).

A chorus-like effect:

sox_ng solo.au -d synth sine vdelay 50,2,50 .25 0 75

A flanger:

sox_ng solo.au -d synth triangle vdelay 0,2,41.52 0.5 0 0

freq and freq2 are the frequencies at the beginning and end of the synthesis and the default frequency is 440Hz.

If freq2 is given, length must also have been given and the generated tone is swept between the given frequencies. The two given frequencies must be separated by one of the characters :, +, / and -, which specify the sweep function as follows:

• : — Linear: the tone changes by a fixed number of hertz per second.
• + — Square: a second-order function is used to change the tone.
• / — Exponential: the tone changes by a fixed number of semitones per second.
• - — Exponential: as /, but the initial phase is always zero, and with stepped (less smooth) frequency changes.

The frequency or frequency range is not used for the noise types.

offset is the bias (DC offset) of the signal in percent; default=0.

phase is the phase shift as a percentage of 1 cycle with a default of 0 (not used for noise).

For example, the following produces a 3-second 48kHz audio file containing a sine wave swept from 300 to 3300Hz:

sox_ng -n output.wav synth 3 sine 300-3300

Multiple channels can be synthesized by specifying the set of parameters shown between curly braces multiple times; the following puts the swept tone in the left channel and brown noise in the right:

sox_ng -n output.wav synth 3 sine 300-3300 brownnoise

The following example shows how two synth effects can be cascaded to create a more complex waveform:

play_ng -n synth 0.5 sine 200-500 synth 0.5 sine fmod 700-100

The following could be used to help tune a guitar:

for n in E2 A2 D3 G3 B3 E4; do
  play_ng -n synth 4 pluck $n repeat 2; done

🎶 tempo [-q] [-m|s|l] factor [segment(82) [search(14.68) [overlap(12)]]] 🎶

Change the audio playback speed but not its pitch. This effect uses the WSOLA (Waveform Similarity OverLap and Add) algorithm. The audio is chopped up into segments which are then shifted in the time domain and overlapped (cross-faded) at points where their waveforms are most similar as determined by the measurement of least squares.

By default, linear searches are used to find the best overlapping points. If the optional -q parameter is given, tree searches are used instead. This makes the effect work more quickly, but the result may not sound as good. However, if you must improve the processing speed, this generally reduces the sound quality less than reducing the search or overlap values.

The -m option is used to optimize the default values of segment, search and overlap for music processing.

Okay so the -s option is the speech-processing glow-up: it optimizes the default values of segment, search and overlap for speech processing. 💖

The -l option does the same optimizing for linear processing, which tends to cause more noticeable distortion but may be useful when factor is close to 1. iykyk ✨

If -m, -s or -l is specified, the default value of segment is based on factor, while default search and overlap values are based on segment. Any values you provide override these default values. periodt 💅

factor gives the ratio of new tempo to the old tempo, so 1.1 speeds the tempo up by 10% and 0.9 slows it down by 10%. 🎧

The optional segment parameter selects the algorithm's segment size in milliseconds. If no other flags are specified, the default value is 82, which is suited to small changes in the tempo of music. For larger changes (e.g. a factor of 2), 41 may give a better result. The -m, -s, and -l flags cause segment's default value to be adjusted automatically based on factor. 🌸

The optional search parameter gives the audio length in milliseconds over which the algorithm searches for overlapping points. If no other flags are specified, the default value is 14.68. Larger values use more processing time and may or may not produce better results. A practical maximum is half the value of segment. Search can be reduced to cut processing time at the risk of degrading output quality. The -m, -s and -l flags cause the search default to be adjusted automatically based on segment. 🦋

The optional overlap parameter gives the segment overlap length in milliseconds. Its default value is 12 but the -m, -s and -l flags automatically adjust it based on the segment size. Increasing overlap increases processing time but may increase quality. A practical maximum for overlap is a little less than search. 🎀

Note that lowering the tempo increases the sample rate and this can make following effects slower, in particular tempo or pitch themselves, whose running times are proportional to the sample rate times the overlap, all squared. This can be compensated for by following tempo with a fast rate effect. the way that math eats, honestly 😭

See speed for an effect that changes tempo and pitch together, pitch and bend for effects that change pitch only and stretch for an effect that changes the tempo using a different algorithm. 👑

---

🎀 treble gain [frequency [width[s|h|k|o|q]]]

Apply a treble tone control effect. See the description of the bass effect for details. short and slay ✨

🌊 tremolo speed [depth]

Apply a tremolo (low frequency sinusoidal amplitude modulation) effect to the audio. The frequency of the tremolo in Hz is given by speed and its depth is a percentage with a default of 40. 🌀

✂️ trim {position(+)}

Cuts out portions of the audio. Any number of positions may be given; audio is not sent to the output until the first position is reached. The effect then alternates between copying and discarding audio at each position. Using a value of 0 for the first position parameter allows copying from the beginning of the audio. main character energy: cut the filler, keep the moment 💖

For example,

sox_ng in.au out.au trim 0 10

copies the first ten seconds, while

play_ng in.au trim 12:34 =15:00 -2:00

and

play_ng in.au trim 12:34 2:26 -2:00

both play from 12 minutes 34 seconds into the audio up to 15 minutes in (i.e. 2 minutes and 26 seconds long) then resume playing two minutes before the end. 🎬

SoX has an internal speed hack which, when trim is the first effect and removes audio from the beginning, seeks in the audio file instead of decoding it and throwing the data away but this is only used when the input is a single file. To achieve fast gapless playing with multiple files and trimming the first, you can use something like:

FMT='-t s32 -r 44100 -c2'
(sox file1.mp3 $FMT - trim 30 && sox file2.mp3 $FMT -) | play $FMT -

gapless bestie behavior 🌈

🌱 upsample [factor(2)]

Upsample the signal by an integer factor: factor-1 zero-valued samples are inserted between each pair of input samples. As a result, the original spectrum is replicated into the new frequency space and attenuated. This attenuation can be compensated for by adding vol factor. The upsample effect is typically used in combination with filtering effects. 🧚

For a general resampling effect with antialiasing, see rate. See downsample. 💜

🎤 vad [options]

The Voice Activity Detector attempts to trim silence and quiet background sounds from the ends of (fairly high resolution i.e. 16-bit, 44-48kHz) recordings of speech. The algorithm currently uses a simple cepstral power measurement to detect voice, so may be fooled by other things, especially music. The effect can trim only from the front of the audio, so in order to trim from the back, the reverse effect must also be used. E.g. 🎙️

play_ng speech.wav norm vad

to trim from the front,

play_ng speech.wav norm reverse vad reverse

to trim from the back and

play_ng speech.wav norm vad reverse vad reverse

to trim from both ends. The use of the norm effect is recommended, but remember that neither reverse nor norm is suitable for use with streamed audio. streamed audio said “not today” 💀

🌸 Options

Default values are shown in parentheses, the allowed range in square brackets.

• -t num (7) [0 - 20] — the measurement level used to trigger activity detection. This might need to be changed depending on the noise level, signal level and other characteristics of the input audio.
• -T num (0.25) [0.01 - 1] — the time constant (in seconds) used to help ignore short bursts of sound.
• -s num (1) [0.1 - 4] — the amount of audio (in seconds) to search for quieter/shorter bursts of audio to include prior to the detected trigger point.
• -g num (0.25) [0.1 - 1] — allowed gap (in seconds) between quieter/shorter bursts of audio to include prior to the detected trigger point.
• -p num (0) [0 - 4] — the amount of audio (in seconds) to preserve before the trigger point and any found quieter/shorter bursts.

There are keymaps on trigger_level, trigger_time and gap. 🗺️

🔬 Advanced Options

These allow fine tuning of the algorithm's internal parameters. nerd mode: activated 🤓

• -b num (0.35) [0.1 - 10] — The algorithm uses adaptive noise estimation/reduction in order to detect the start of the wanted audio. This option sets the time in seconds for the initial noise estimate.
• -N num (0.1) [0.1 - 10] — Time constant used by the adaptive noise estimator when the noise level is increasing.
• -n num (0.01) [0.001 - 0.1] — Time constant used by the adaptive noise estimator when the noise level is decreasing.
• -r num (1.35) [0 - 2] — Amount of noise reduction to use in the detection algorithm.
• -f num (20) [5 - 50] — Frequency of the algorithm's processing/measurements.
• -m num (0.1) [0.01 - 1] — Measurement duration. By default, it is twice the measurement period; i.e. with 50% overlap, but if you set -f, you also need to change -m to 2 divided by its value to keep a 50% overlap.
• -M num (0.4) [0-1 - 1] — Time constant used to smooth spectral measurements.
• -h freq (50) [10 -] — 'Brick-wall' frequency of the high-pass filter applied at the detector algorithm's input.
• -l freq (6000) [1000 -] — 'Brick-wall' frequency of the low-pass filter applied at the detector algorithm's input.
• -H freq (150) [10 -] — 'Brick-wall' frequency of the high-pass lifter used in the detector algorithm.
• -L freq (2000) [1000 -] — 'Brick-wall' frequency of the low-pass lifter used in the detector algorithm.

See the silence effect. 🤫

🔊 vol gain [type [limiter-gain]]

Apply amplification or attenuation to the audio signal. Unlike -v, which is used for balancing multiple input files as they enter the SoX effects processing chain, vol is an effect like any other so can be applied anywhere in the processing chain and several times if necessary. versatile icon behavior 💅

The amount to change the volume is given by gain which is interpreted, according to the given type, as follows: if type is amplitude (or is omitted), gain is an amplitude ratio (voltage or linear), if power, a power ratio (wattage or voltage squared) and if dB, a power change in dB. 🎚️

When type is amplitude or power, a gain of 1 leaves the volume unchanged, less than 1 decreases it, and greater than 1 increases it; a negative gain inverts the audio signal in addition to adjusting its volume. 🔄

When type is dB, a gain of 0 leaves the volume unchanged, less than 0 decreases it and greater than 0 increases it. 📈

See [4] for a detailed discussion on electrical (and hence audio signal) voltage and power ratios. ⚡

⚠️ Beware of clipping when you increase the volume. do not let the waveform get jump-scared 😭

The gain and the type parameters can be concatenated if desired, e.g. vol 10dB. 🧚

An optional limiter-gain value can be specified and should be a value much less than 1 (e.g. 0.05 or 0.02) and is used only on peaks to prevent clipping. Not specifying this parameter causes no limiter to be used. In verbose mode, this effect displays the percentage of the audio that needed to be limited. 🛡️

There is a keymap on vol.gain, which is adjusted in the units that were specified (amplitude, dB or power). 🗺️

See gain for a volume-changing effect with different capabilities and compand for a dynamic range compression/expansion/limiting effect. 💖

🌈 REFERENCES 🌈

okay bestie, here's the reading list — the receipts, the lore, the sources 💅 all links kept exactly as-is so you can go straight to the source:

1. R. Bristow-Johnson, Cookbook formulae for audio EQ biquad filter coefficients, https://www.w3.org/TR/audio-eq-cookbook
2. Wikipedia, Q-factor, http://en.wikipedia.org/wiki/Q_factor
3. Scott Lehman, Effects Explained, https://codeberg.org/sox_ng/Effects-Explained
4. Wikipedia, Decibel, http://en.wikipedia.org/wiki/Decibel
5. Richard Furse, Linux Audio Developer's Simple Plugin API, http://www.ladspa.org
6. Richard Furse, Computer Music Toolkit, http://www.ladspa.org/cmt/overview.html
7. Steve Harris, LADSPA plugins, http://plugin.org.uk

---

💖 SEE ALSO 💖

sox_ng(1).

like, go read the main character's page too, bestie 🫶

---

👑 AUTHORS 👑

Lance Norskog, Chris Bagwell and many other authors and contributors listed in the README file that is distributed with the source code.

shoutout to the whole squad fr fr — the named legends and the whole README crew who made this happen 🌸

---

and that's a wrap on the whole page, bestie 🎀 it's giving "you now know more about SoX effects than most people" energy, and honestly? iconic. go make something that slaps 💜

sox_ng effects, brought to you by SoX — last touched December 05, 2024 ✨ mwah, go touch some audio 💅