taha-firoz / kaldi2scribe

A quick and dirty way to run your Kaldi model

kaldi2scribe

This repo represents weeks of frustration and difficulty I faced while trying to run my trained Kaldi model. Yes, I aware of the fact that Kaldi is a tool that was developed for researchers but the undecipherable documentation is not justified.

Unless you have a deep grasp of how Kaldi and ASR in general works you'll be getting nowhere. If you've been in a similar situation where you've followed the notorious Kaldi for Dummies tutorial and finally want to reap the fruits of your labor and see audio being transcribed, you'd be out of luck. It's as if no one from Kaldi's team every wondered whether anyone would want to run their models and see it transcribe text. You'll find excellent resources like PyKaldi Examples or Nithing Rao's Medium article but again for someone who's just following Kaldi for Dummies they probably won't understand how to get them to work. Also given that a majority of the examples aren't for simple GMM setups but are based on running more complex TDNN or NN acoustic models, they aren't going to help the Kaldi for Dummies gang.

So I've made this small tool that takes in audio files, your model, and kaldi and spits out their transcriptions. I used Anwarvic's really well written Arabic-Speech-Recognition repo to guide me through writing it. It is basically his work that I've gutted to just spit out transcriptions instead of evaluating your model's accuracy.

Install

PyKaldi is available on conda, unless you plan on building it from scratch I'd highly recommend you use conda.

$ conda install --yes --file requirements.txt

You also need a copy of kaldi that's been built, if you've already trained a model then I guess you'll have a copy lying around. Just place the files this way:

kaldi2scribe.py
|
model root
└───exp
│   └───tri1
|   |   ...
kaldi root
└───nightmare fuel
|   |   ...

Tbh it doesn't really matter where everything is just make sure you CWD is at the same level as the Kaldi's root and yeah that's about it.

Usage

$ python3 kaldi2scribe.py --modelroot '/wsj/s5' --kaldiroot '/kaldi' --type 'tri1' --input '/transcription_test'

Help

$ python3 kaldi2scribe.py -h

Note

This is a very basic tool meant to help complete kaldi beginners to test their models, if you're looking at it for anything more then you're out of luck. If you have a better grasp on kaldi then make sure to contribute, specially to add support for more models.