SPHINX_LM_EVAL(1)

✨ SPHINX_LM_EVAL ✨

ok so bestie, this one lives in the General Commands Manual section and it's called sphinx_lm_eval, and it's here to do exactly one iconic thing: tell you how perplexed a transcription is 💅

---

🌸 NAME 🌸

sphinx_lm_eval — evaluate perplexity of a transcription

---

🦋 SYNOPSIS 🦋

sphinx_lm_eval [ options ]...

---

💖 DESCRIPTION 💖

So here's the vibe: this program evaluates the perplexity of a text file according to a given language model. That's the whole main character moment. 🔥

And the text file? It's assumed to be in transcript format, meaning one utterance per line, delimited by <s> and </s>. Like, every single line is its own little utterance, wrapped up all cozy. No cap.

---

🎀 the options (yes, all of them, we're not skipping a single one) 🎀

• -help — Shows the usage of the tool. that's it, that's the flag, it just tells you how to use it 🫶
• -lm model file — the model file, bestie. your language model lives here.
• -lmctlfn — file listing a set of language models. it's giving "control file" energy. iykyk.
• -lmname — of language model in -lmctlfn to use for all utterances. literally every single utterance gets this one.
• -logbase — Base in which all log-likelihoods calculated. pick your base, it's your world ✨
• -lsn file to evaluate — this is the file you're gonna evaluate. the moment of truth.
• -lw — Language model weight. how much say does the LM get? your call.
• -mmap — Use memory-mapped I/O for reading binary LM files. fancy. fast. obsessed.
• -probdef — definition file for classes in LM. classes gotta be defined somewhere, so.
• -text string to evaluate — string to evaluate. yes, you can just hand it a string directly instead of a whole file, slay.
• -uw — Unigram probability weight (interpolated with uniform distribution). it's getting blended in there with the uniform distribution, periodt.
• -verbose — Print details of perplexity calculation. turn this on if you wanna see everything the math is doing. not me being nosy about my own perplexity calc 😭
• -wip — Word insertion probability. that's the whole thing. the words, the inserting, the probability. 💫

---

👑 AUTHOR 👑

David Huggins-Daines <dhdaines@gmail.com>

---

🌈 COPYRIGHT 🌈

Copyright © 2007-2008 Carnegie Mellon University. See the file COPYING included with this package for more information.

---

ok and one last thing before you go — this whole page is from the 2008-05-12 era, so shoutout to that date for keeping it real all these years 🩷 go evaluate some perplexity bestie, you got this 💜✨