awesome-kaldi

The most basic and comprehensive article about the creation of Language-Models

A really good presentation by Andrew Maas for better understanding the GMM-based phoneme alignment.

A good small course (~3 hours) from Nanyang technological university that covers the idea of WFSTs in a really straight forward and visual way.

The "bible" for understanding WFST-based systems for Speech recognition.

The HTK book is for another ASR toolkit but it highlihts the basics of speech recognition in a a really intuitive and graphic way.

The article that describes the usage of TDNNs in Kaldi

an article about the BLSTM basic recipe in Kaldi.

The original article that described Kaldi and the different parts of the project. It should be noted that some parts of that article are outdated.

This presentation is extremely long but also extremely helpful. Its the most complete source of information about the training process and its development.

A good in depth tutorial about the training process with a lot of code examples.

A Medium post (by me) regarding the general structure of the Kaldi project and its different parts.

A tutorial by Josh Meyer for specifically running Kaldi with DNN

The basic tutorial in the Kaldi documentation. It is really good for "hands on" experience but it is not so well explained.

A tutorial about X-Vectors and Speaker Diarization.

If you have multiple datasets and you want to combine all of the manually, there is no need to do it file after file. this script will take an entire data directory and will combine all the files into the same new directory.

If you don't have a lot of data You can always train a Kaldi model from the closest domain to your domain and then take the

A project that helps transferring the Kaldi model into ONNX so you could easily use the model in different frameworks.

this script will help you to change the speaking speed of different utterances without creating excess files. It does this by implementing an SoX command to your wav file and copying and editing all the other files in your folder. Using this script and also the next one is a must-have in most state-of-the-art systems and will help your model to generalize better.

this script will do exactly the same but will change the volume of the utterances.

You want to make a new model for different sampling rate but you don't want to manually re-sample you entire data? this script will help you to do it, again with a SoX command.

You can also compile the Kaldi project in a way that will work directly on android devices. That might not be a good idea with a heavy model, but can be used to more constrained models.

A tool that helps to adapt nnet3 chain models to a different language model.

this is a nice project that will help you to integrate Kaldi toolkit and the

A good example for a project that handles both training and decoding. It is being build for Estonian but can be easily transformed into any language.

A new excutable that was

A framework that combines TensorFlow and Kaldi in the context of speaker verification/identification tasks. The project has some pretrained model that were trained on huge datasets.

If you want to understand the different parts of the Decoding graph you should probably read this. It is required to understand those concepts for debugging your graph in the development of a new model.

Some advanced notes that is highly recommended to read if you want to be a more trained user.

The models trained on Kaldi website. In

Trained on 700+- hours of german data. Gets updated from time to time.