wav2letter: FAIR’s automatic speech recognition toolkit

综合技术 2018-01-01 阅读原文


wav2letter is a simple and efficient end-to-end Automatic Speech Recognition (ASR) system from Facebook AI Research. The original authors of this implementation are Ronan Collobert, Christian Puhrsch, Gabriel Synnaeve, Neil Zeghidour, and Vitaliy Liptchinsky.

wav2letter implements the architecture proposed in Wav2Letter: an End-to-End ConvNet-based Speech Recognition System
and Letter-Based Speech Recognition with Gated ConvNets

If you want to get started transcribing speech right away, we providepre-trained modelsfor the Librispeech


Our approach is detailed in two scientific contributions:

  author    = {Ronan Collobert and Christian Puhrsch and Gabriel Synnaeve},
  title     = {Wav2Letter: an End-to-End ConvNet-based Speech Recognition System},
  journal   = {CoRR},
  volume    = {abs/1609.03193},
  year      = {2016},
  url       = {http://arxiv.org/abs/1609.03193},


  author    = {Vitaliy Liptchinsky and Gabriel Synnaeve and Ronan Collobert},
  title     = {Letter-Based Speech Recognition with Gated ConvNets},
  journal   = {CoRR},
  volume    = {abs/1712.09444},
  year      = {2017},
  url       = {http://arxiv.org/abs/1712.09444},

If you use wav2letter or related pre-trained models, then please cite one of these papers.


  • A computer running MacOS or Linux.
  • Torch
    . We detail in the following how to install it.
  • For training on CPU: Intel MKL
  • For training on GPU: NVIDIA CUDA Toolkit (cuDNN v5.1 for CUDA 8.0)
  • For reading of audio file: Libsndfile
    - should be available in any standard distribution.
  • For standard speech features: FFTW
    - should be available in any standard distribution.



If you plan to train on CPU, it is highly recommended to install Intel MKL

Update your .bashrc file with the following:

# We assume Torch will be installed in $HOME/usr.
# Change according to your needs.
export PATH=$HOME/usr/bin:$PATH

# This is to detect MKL during compilation
# but also to make sure it is found at runtime.

if [ ! -d "$INTEL_DIR" ]; then
    echo "$ warning: INTEL_DIR out of date"
if [ ! -d "$MKL_DIR" ]; then
    echo "$ warning: MKL_DIR out of date"
if [ ! -d "$MKL_INC_DIR" ]; then
    echo "$ warning: MKL_INC_DIR out of date"

# Make sure MKL can be found by Torch.

LuaJIT + LuaRocks

The following installs luaJIT and luarocks locally in $HOME/usr
. If you want a system-wide installation, remove the -DCMAKE_INSTALL_PREFIX=$HOME/usr

git clone https://github.com/torch/luajit-rocks.git
cd luajit-rocks
mkdir build; cd build
make -j 4
make install
cd ../..

In the next sections, we assume luarocks
and luajit
are in $PATH
. If they are not - and assuming you installed them locally in $HOME/usr
- you can instead run ~/usr/bin/luarocks
and ~/usr/bin/luajit

KenLM Language Model Toolkit

If you plan to use the wav2letter decoder, you will need KenLM.

KenLM requires Boost

# make sure boost is installed (with system/thread/test modules)
# actual command might vary depending on your system
sudo apt-get install libboost-dev libboost-system-dev libboost-thread-dev libboost-test-dev

Once boost is properly installed, you may install KenLM:

wget https://kheafield.com/code/kenlm.tar.gz
tar xfvz kenlm.tar.gz
cd kenlm
mkdir build && cd build
make -j 4
make install
cp -a lib/* ~/usr/lib # libs are not installed by default 🙁
cd ../..

and TorchMPI

If you plan to use multi-CPU/GPUs (and/or multi-machines), you will need OpenMPI and TorchMPI.

Disclaimer: it is highly encouraged to recompile OpenMPI yourself. OpenMPI binaries on standard distributions come with a lot of variance in the compilation flags. Certain flags are crucial to successfully compile and run TorchMPI.

First install OpenMPI:

wget https://www.open-mpi.org/software/ompi/v2.1/downloads/openmpi-2.1.2.tar.bz2
tar xfj openmpi-2.1.2.tar.bz2
cd openmpi-2.1.2; mkdir build; cd build
./configure --prefix=$HOME/usr --enable-mpi-cxx --enable-shared --with-slurm --enable-mpi-thread-multiple --enable-mpi-ext=affinity,cuda --with-cuda=/public/apps/cuda/9.0
make -j 20 all
make install

Note: works the same with openmpi-3.0.0.tar.bz2, but --enable-mpi-thread-multiple needs then to be removed.

You may now install TorchMPI:

MPI_CXX_COMPILER=$HOME/usr/bin/mpicxx ~/usr/bin/luarocks install torchmpi

Torch and other Torch packages

luarocks install torch
luarocks install cudnn # for GPU support
luarocks install cunn # for GPU support

wav2letter packages

git clone https://github.com/facebookresearch/wav2letter.git
cd wav2letter
cd gtn && luarocks make rocks/gtn-scm-1.rockspec && cd ..
cd speech && luarocks make rocks/speech-scm-1.rockspec && cd ..
cd torchnet-optim && luarocks make rocks/torchnet-optim-scm-1.rockspec && cd ..
cd wav2letter && luarocks make rocks/wav2letter-scm-1.rockspec && cd ..
# Assuming here you got KenLM in $HOME/kenlm
# And only if you plan to use the decoder:
cd beamer && KENLM_INC=$HOME/kenlm luarocks make rocks/beamer-scm-1.rockspec && cd ..

Training wav2letter models

Data pre-processing

The data
folder contains a number of scripts for preprocessing various datasets. For now we provide only LibriSpeech and TIMIT.

Below is an example on how to preprocess LibriSpeech ASR corpus:

wget http://www.openslr.org/resources/12/dev-clean.tar.gz
tar xfvz dev-clean.tar.gz
# repeat for train-clean-100, train-clean-360, train-other-500, dev-other, test-clean, test-other
luajit ~/wav2letter/data/librispeech/create.lua ~/LibriSpeech ~/librispeech-proc
luajit ~/wav2letter/data/utils/create-sz.lua librispeech-proc/train-clean-100 librispeech-proc/train-clean-360 librispeech-proc/train-other-500 librispeech-proc/dev-clean librispeech-proc/dev-other librispeech-proc/test-clean librispeech-proc/test-other


mkdir experiments
luajit ~/wav2letter/train.lua --train -rundir ~/experiments -runname hello_librispeech -arch ~/wav2letter/arch/librispeech-glu-highdropout -lr 0.1 -lrcrit 0.0005 -gpu 1 -linseg 1 -linlr 0 -linlrcrit 0.005 -onorm target -nthread 6 -dictdir ~/librispeech-proc  -datadir ~/librispeech-proc -train train-clean-100+train-clean-360+train-other-500 -valid dev-clean+dev-other -test test-clean+test-other -gpu 1 -sqnorm -mfsc -melfloor 1 -surround "|" -replabel 2 -progress -wnorm -normclamp 0.2 -momentum 0.9 -weightdecay 1e-05

Training on multiple GPUs

Use OpenMPI to spawn multiple training processes, one per GPU:

mpirun -n 2 --bind-to none  ~/TorchMPI/scripts/wrap.sh luajit ~/wav2letter/train.lua --train -mpi -gpu 1 ...

We assume here mpirun
is in $PATH

Running the decoder (inference)

We need to do few pre-processing steps to run the decoder.

We first create a dictionary of letters, which includes the special repetition letters we use in wav2letter:

cat ~/librispeech-proc/letters.lst >> ~/librispeech-proc/letters-rep.lst && echo "1" >> ~/librispeech-proc/letters-rep.lst && echo "2" >> ~/librispeech-proc/letters-rep.lst

We then get a language model, and pre-process it. Here, we will use the pre-trained language models for LibriSpeech
, but one can also train its own with KenLM. We then pre-process it to transform words in low caps, and produce their letter transcriptions with the repetition letters in a particular dictionary dict.lst
. The script might warn you about words which are incorrectly transcribed, due to insufficient number of repetitions letters (here 2, with -r 2
). This is not a problem in our case, as these words are rare.

wget http://www.openslr.org/resources/11/3-gram.pruned.3e-7.arpa.gz luajit
~/wav2letter/data/utils/convert-arpa.lua ~/3-gram.pruned.3e-7.arpa.gz ~/3-gram.pruned.3e-7.arpa ~/dict.lst -preprocess ~/wav2letter/data/librispeech/preprocess.lua -r 2 -letters letters-rep.lst

Note: one can use the pre-trained 4-gram language model 4-gram.arpa.gz
instead; pre-processing will take longer.

Optional: subsequent loading of the language model can be made faster by converting it to a binary format with KenLM (we assume here KenLM is in your $PATH

build_binary 3-gram.pruned.3e-7.arpa 3-gram.pruned.3e-7.bin

We can now generate emissions for a particular trained model, running test.lua
on a dataset. The script also displays Letter Error Rate (LER) and Word Error Rate (WER) - the latter being computed with no post-processing of the acoustic model.

luajit ~/wav2letter/test.lua ~/experiments/hello_librispeech/001_model_dev-clean.bin -progress -show -test dev-clean -save

Once the emissions are stored, the decoder can be ran to compute the WER obtained by constraining the decoding with a particular language model:

luajit ~/wav2letter/decode.lua ~/experiments/hello_librispeech dev-clean -show -letters ~/librispeech-proc/letters-rep.lst  -words ~/dict.lst -lm ~/3-gram.pruned.3e-7.arpa -lmweight 3.1639 -beamsize 25000 -beamscore 40 -nthread 10 -smearing max -show

Pre-trained models

We provide a fully pre-trained model for LibriSpeech:

wget https://s3.amazonaws.com/wav2letter/models/librispeech-glu-highdropout.bin

To transcribe speech using this model, you need to follow the some of therequirements,installation, anddecodingparts of this README.

NOTE: the model was pre-trained on Facebook infrastructure, so you need to run test.lua
with slightly different parameters to use it:

luajit ~/wav2letter/test.lua ~/librispeech-glu-highdropout.bin -progress -show -test dev-clean -save -datadir ~/librispeech-proc/ -dictdir ~/librispeech-proc/ -gfsai

Join the wav2letter community

for how to help out.


as well as the PATENTS




用好Lua+Unity,让性能飞起来—LuaJIT性能坑详解... 导语:大家都知道LuaJIT比原生Lua快,快在JIT这三个字上。但实际情况是,LuaJIT的行为十分复杂。尤其JIT并不是一个简单的把代码翻译成机器码的机制,背后有很多会影响性能的因素存在,下面笔者将带大家一一说明。 这是侑虎科技的原创文章,感谢作者招文勇供稿,欢迎转发分享,未经作者...
Code reading: LuaJIT I have started reading LuaJIT sources. I like the fact that the source code is compact and it is reasonable to print and read a whole file (or read i...
Speech Recognition : Connectionist Temporal Classi... Welcome to the deep learning in speech recognition series. This is the second part in three part. In the first part, we discussed how to represent aud...
Nginx + LuaJIT + Webhook Install LuaJIT ngx devel kit lua-nginx-module Nginx Base apt install make gcc git libpcre3 libpcre3-dev zlib1g-dev libssl-dev ...
Speech Recognition : Decoding Welcome to the deep learning in speech recognition series. This is the third part in three part. In the first part, we discussed how to represent audi...