5 ms·

Kaldi Speech Recognition Toolkit

71 points by based2 10y ago

lobius 10y ago
Still based on MFCC features, which are very noisy / high entropy.
- gok 10y ago
  This is completely wrong. Kaldi has code to compute all kinds of features (filter bank, pitch, PLP...) and the many recipes use them.
  - lobius 10y ago
    But not spectral peaks, which is what audio fingerprinting services like Shazam use (very successfully too it seems)
    - skoocda 10y ago
      There's been very little use of spectral peaks in LVCSR for the past ~10 years.
    - unlikelymordant 10y ago
      Shazam and asr are very different problems. The things that work for one don't work for the other.
blueyes 10y ago
This used to be the basis of Siri I think. Not sure it is any more, and not sure it's state of the art of speech recognition, given what neural nets can do.
- gok 10y ago
  Kaldi contains implementations of state of the art ASR techniques, including many neural network approaches.
- galv 10y ago
  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 http://www.danielpovey.com/publications.html He is the main maintainer of the project.
jomamaxx 10y ago
Does anyone have any benchmarks for this?
- galv 10y ago
  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/s5/RESULTS 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.
- skoocda 10y ago
  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 https://github.com/syhw/wer_are_we
redgetan 10y ago
What's a good resource to learn more about speaker diarization so that I can learn how to use existing tools properly (tweaking and modifying them according to my needs such as improving accuracy). Some diarizers that Im interested in trying so far are - PyCASP, pyAudioAnalysis, dlg-segmenter
- nshm 10y ago
  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.
braindead_in 10y ago
We have an open position for an Kaldi Engineer, in case anyone's interested. https://news.ycombinator.com/item?id=12630618 https://news.ycombinator.com/item?id=12630618
cnorthwood 10y ago
Kaldi is used in production at the BBC: http://bbcnewslabs.co.uk/projects/transcriptor/ http://bbcnewslabs.co.uk/projects/transcriptor/
skykooler 10y ago
Is there a good resource for getting started with Kaldi? I wanted to test it as an alternative to Julius, but all the documentation seems to assume you have a doctorate in speech recognition technologies or something.
- galv 10y ago
  This guide was written by a user who thought the learning curve was too high: http://kaldi-asr.org/doc/kaldi_for_dummies.html 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 http://mi.eng.cam.ac.uk/~mjfg/mjfg_NOW.pdf - The finite state transducers paper http://www.cs.nyu.edu/~mohri/pub/hbka.pdf http://www.cs.nyu.edu/~mohri/pub/hbka.pdf
  - skoocda 10y ago
    To add to this, Daniel Povey's lectures [0] are fairly useful (although not exactly colourful, per se). Would definitely second the FST paper. You'll get shunned from the forums if you haven't read that one. [0] http://www.danielpovey.com/kaldi-lectures.html http://www.danielpovey.com/kaldi-lectures.html

71 points by based2 10y ago

lobius 10y ago

Still based on MFCC features, which are very noisy / high entropy.

gok 10y ago
This is completely wrong. Kaldi has code to compute all kinds of features (filter bank, pitch, PLP...) and the many recipes use them.
- lobius 10y ago
  But not spectral peaks, which is what audio fingerprinting services like Shazam use (very successfully too it seems)
  - skoocda 10y ago
    There's been very little use of spectral peaks in LVCSR for the past ~10 years.
  - unlikelymordant 10y ago
    Shazam and asr are very different problems. The things that work for one don't work for the other.

blueyes 10y ago

This used to be the basis of Siri I think. Not sure it is any more, and not sure it's state of the art of speech recognition, given what neural nets can do.

gok 10y ago
Kaldi contains implementations of state of the art ASR techniques, including many neural network approaches.
galv 10y ago
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 http://www.danielpovey.com/publications.html He is the main maintainer of the project.

jomamaxx 10y ago

Does anyone have any benchmarks for this?

galv 10y ago
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/s5/RESULTS 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.
skoocda 10y ago
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 https://github.com/syhw/wer_are_we

redgetan 10y ago

What's a good resource to learn more about speaker diarization so that I can learn how to use existing tools properly (tweaking and modifying them according to my needs such as improving accuracy). Some diarizers that Im interested in trying so far are - PyCASP, pyAudioAnalysis, dlg-segmenter

nshm 10y ago
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.

braindead_in 10y ago

We have an open position for an Kaldi Engineer, in case anyone's interested. https://news.ycombinator.com/item?id=12630618 https://news.ycombinator.com/item?id=12630618

cnorthwood 10y ago

Kaldi is used in production at the BBC: http://bbcnewslabs.co.uk/projects/transcriptor/ http://bbcnewslabs.co.uk/projects/transcriptor/

skykooler 10y ago

Is there a good resource for getting started with Kaldi? I wanted to test it as an alternative to Julius, but all the documentation seems to assume you have a doctorate in speech recognition technologies or something.

galv 10y ago
This guide was written by a user who thought the learning curve was too high: http://kaldi-asr.org/doc/kaldi_for_dummies.html 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 http://mi.eng.cam.ac.uk/~mjfg/mjfg_NOW.pdf - The finite state transducers paper http://www.cs.nyu.edu/~mohri/pub/hbka.pdf http://www.cs.nyu.edu/~mohri/pub/hbka.pdf
- skoocda 10y ago
  To add to this, Daniel Povey's lectures [0] are fairly useful (although not exactly colourful, per se). Would definitely second the FST paper. You'll get shunned from the forums if you haven't read that one. [0] http://www.danielpovey.com/kaldi-lectures.html http://www.danielpovey.com/kaldi-lectures.html