galvASR: Daniel Galvez's Automatic Speech Recognition

Warning

This library is under active development. There will definitely be bugs.

Motivation

I am one of the maintainers of the Kaldi Speech Recognition system. It is, in a lot of people's opinions, probably the best free open source speech recognition system. Its style of speech recognition models, hybrid HMM-DNNs, are at the time of writing the state-of-the-art in speech recognition, thanks to innovations like time-delay neural networks and lattice-free MMI. No other software library supports as many datasets as it does out-of-the-box via its recipe ("egs") scripts. One of these libraries in particular, Librispeech, was developed by the team as the first large (~1000+ hours) free speech recognition corpus. I know a few too many speech startups that got their start based solely on this dataset!

Lately, we've seen a rise in deep learning frameworks, and with that an interest in "end-to-end" or fully-neural speech recognition systems lately.

Basically all of these frameworks have chosen a python-based interface, with a C++ core. This is a marked step away from Kaldi's core of C++ executables that are driven by bash scripts. In my interactions with people in industry, it's become clear that many people would like a more familiar Python interface, while still having many of the advantages of Kaldi.

Therefore, galvASR provides tensorflow datasets ops caffe2 ops to read Kaldi's flexible key-value pair table files. This means you can take any Kaldi-prepared data directory generated by one of Kaldi's existing recipes and do training on it immediately with your framework of choice.

I've also noticed that there has been a lot of fragementation recently in terms of making work reproducible in the field of speech recognition, with many research groups seeking improvement via very large datasets. However, none of these groups has studied whether other group's models get the same benefits from larger datasets. I would like to test the hypothesis that there may be little significant difference between these models on the same datasets.

Therefore, another goal of galvASR is to make a large, multidataset benchmarks, one using only free data (we can get ~2000 hours this way, the size of the common Switchboard-Fisher benchmark), and another combining free and non-free datasets (I'd like to reach 10,000 hours).

Install

See the Dockerfile to get the dependencies.

Note that you will need nvidia-docker to run this because of the CUDA dependency.

cd scripts/docker
docker build -t galv-asr DefaultBuild

Build

This library uses submodules! Read more about how to handle those here

git clone --recursive https://github.com/galv/galvASR.git
cd galvASR &&
mkdir build &&
cd build &&
cmake -DPYTHON_INSTALL_EDITABLE=NO -DCMAKE_INSTALL_PREFIX=../install -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DUSE_CAFFE2=NO -DUSE_TENSORFLOW=YES ..
make -k -j27

Run