Need help with vosk?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

231 Stars 27 Forks Apache License 2.0 15 Commits 3 Opened issues


VOSK Speech Recognition Toolkit

Services available


Need anything else?

Contributors list

# 7,297
15 commits

For Kaldi API for Android and Linux please see Vosk API. This is a server project.

This is Vosk, the lifelong speech recognition system.


As of 2019, the neural network based speech recognizers are pretty limited in terms of amount of the speech data they can use in training and require enormous computing power and time to train and optimize the parameters. Neural networks have problems with human-like one shot learning, their decisions are not very robust to unseen conditions and hard to understand and correct.

That is why we decided to build a system based on large signal database concept. We apply audio fingerprinting scheme. The audio is segmented on chunks, the chunks are stored in the database based on LSH hash value. During decoding we simply lookup the chunks in the database to get the idea what are the possible phones. That helps us to make a proper decision on decoding results.

The advantages of this approach are:

  • We can quickly train on 100000 hours of speech data on very simple hardware
  • We can easily correct recognizer behavior just by adding samples
  • We can make sure that recognition result is correct because it is sufficiently represented in the training dataset
  • We can parallelize training across thousands of nodes
  • We support lifelong learning paradigm
  • We can use this method together with more common neural network training to improve recognition accuracy
  • The system is robust against noise

The disandvantages are:

  • The index is really huge, it is not expected to fit a memory of single server
  • The generalization capabilities of the model are quite questionable, at the same time the generalization capabilities of the neural networks are also questionable.
  • For now the segmentation requires conventional ASR, but in the future we might segment ourselves.

The nice to have things in the future would be:

  • Multilingual training
  • Our own segmentation
  • The tool to reduce the model to fit the mobile
  • Specialized hardware to implement this AI paradigm


To install the requirements run

pip3 install -r requirements.txt

To prepare the training/verification data create the following two files:

  • wav.scp
    list to map uterances to wav files in filesystem
  • phones.txt
    the CTM file with phonemes and timings. It could be CTM file from the alignment or it could be a CTM file from the decoding

You can create them with Kaldi ASR toolkit


To add the data to the database run

python3 wavs-train.txt phones-train.txt data.idx

That will add the data to the database data.idx or create a new one


To verify decoding results run

python3 wavs-test.txt phones-test.txt data.idx

The tool will search for segments in the index and report suspicious segments which you can additionally check and later add to the database to improve the accuracy of recognition.

Related papers and links

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.