Kaldi Speech Recognition Toolkit

Kaldi contains implementations of state of the art ASR techniques, including many neural network approaches.

It is state of the art in speech recognition exactly because it uses neural networks. I helped develop its nnet3 neural network library and have published based on work I've done on it. No other open source toolkit is better than it right now, both because it has an open license (Apache2) and because it has a larger community than any other speech toolkit.

You can see a lot of state of the art work done using Kaldi at Dan Povey's web page: http://www.danielpovey.com/publications.html He is the main maintainer of the project.

Previous generation technology used GMM and agglomerative clustering. Example is Pycasp, lium diarization. For best performance you need ivector extraction and ivector clustering, that technology is not fully available in open source yet.

This guide was written by a user who thought the learning curve was too high: http://kaldi-asr.org/doc/kaldi_for_dummies.html

The reality is that you will have to study a lot to truly understand the toolkit though. Speech recognition isn't as simple as image recognition where you can just throw a neural network at the problem (that might come off as offensive, but it really is more complicated). Note that you do not need a doctorate in speech recognition to understand it, as I don't have one.

In hindsight, the resources most useful to me when I first began working with Kaldi were:

- the HTK book (behind a login wall, annoyingly) - This book http://mi.eng.cam.ac.uk/~mjfg/mjfg_NOW.pdf - The finite state transducers paper http://www.cs.nyu.edu/~mohri/pub/hbka.pdf

Every dataset has scripts to run reproducible baselines. Look at the RESULTS file for each dataset to see the results of various models. e.g., https://github.com/kaldi-asr/kaldi/blob/master/egs/tedlium/s...

The training of these models can be reproduced on your own machine using a run*.sh script plug-and-play style, but you'll have to do some digging to find what scripts correspond to which models in the RESULTS files.

Since Kaldi is a toolkit, it can be used to build nearly any ASR architecture. See here [0] for a comprehensive comparison of the Word Error Rate of various architectures.

[0]: https://github.com/syhw/wer_are_we

This is completely wrong. Kaldi has code to compute all kinds of features (filter bank, pitch, PLP...) and the many recipes use them.