C++ Python C Shell CMake Makefile Other
Latest commit 7735ac7 Dec 13, 2017 @lissyx lissyx Merge pull request #1088 from wgent/patch-1
[Docs] Renames output_model.pb to output_graph.pb in README

X-DeepSpeech: NOBUILD
Permalink
Failed to load latest commit information.
bin Better resuming; additional README content; small improvements Nov 29, 2017
data Revert "Added language model with apostrophe" Nov 17, 2017
doc Switch docs theme from alabaster to classic Aug 29, 2017
images Compressed gif Nov 28, 2017
native_client Add missing libtensorflow_framework.so Dec 8, 2017
resources Fix #13; Fix #14; Reporting of WERs by index.htm Oct 14, 2016
taskcluster Add OSX testing Dec 7, 2017
util Common Voice importer Nov 23, 2017
.compute Fix #590; Fix #551; Fix #615; Better process exit and exception handlin Jun 7, 2017
.gitattributes Remove current re-scoring of decoder output and switch to custom op Sep 13, 2017
.gitignore Ignore all .egg-infos Nov 17, 2017
.taskcluster.yml Switch to JSON-e fixed version Nov 30, 2017
DeepSpeech.py Update hyperparameters Nov 17, 2017
ISSUE_TEMPLATE.md Create an issue template Nov 27, 2017
LICENSE Added LICENSE Sep 20, 2016
README.md Merge pull request #1088 from wgent/patch-1 Dec 13, 2017
README.website.md fixing small issue with rhetorics Aug 18, 2017
index.htm Fix #328; distributed TensorFlow Mar 31, 2017
requirements.txt Remove scipy 0.19.1 hard-requirement Dec 4, 2017
tc-benchmark-tests.sh TaskCluster enabling of libdeepspeech_model.so Nov 3, 2017
tc-brew-tests.sh Add OSX testing Dec 7, 2017
tc-cpp-ds-tests-prod.sh Add OSX testing Dec 7, 2017
tc-cpp-ds-tests.sh Add OSX testing Dec 7, 2017
tc-node-tests-prod.sh Add OSX testing Dec 7, 2017
tc-node-tests.sh Add OSX testing Dec 7, 2017
tc-python-tests-prod.sh Add OSX testing Dec 7, 2017
tc-python-tests.sh Add OSX testing Dec 7, 2017
tc-schedule.sh Adding build-inhibition handling Nov 28, 2017
tc-tests-utils.sh Add OSX testing Dec 7, 2017
tc-train-tests.sh Run training test against forked and upstream TensorFlow Nov 30, 2017

README.md

Project DeepSpeech

Documentation Status Task Status

Project DeepSpeech is an open source Speech-To-Text engine, using a model trained by machine learning techniques, based on Baidu's Deep Speech research paper. Project DeepSpeech uses Google's TensorFlow project to make the implementation easier.

Usage

Pre-built binaries that can be used for performing inference with a trained model can be installed with pip. Proper setup using virtual environment is recommended and you can find that documented below.

A pre-trained English model is available for use, and can be downloaded using the instructions below.

Once everything is installed you can then use the deepspeech binary to do speech-to-text on short, approximately 5 second, audio files (currently only WAVE files with 16-bit, 16 kHz, mono are supported in the Python client):

pip install deepspeech
deepspeech models/output_graph.pb my_audio_file.wav models/alphabet.txt

Alternatively, quicker inference (The realtime factor on a GeForce GTX 1070 is about 0.44.) can be performed using a supported NVIDIA GPU on Linux. (See the release notes to find which GPU's are supported.) This is done by instead installing the GPU specific package:

pip install deepspeech-gpu
deepspeech models/output_graph.pb my_audio_file.wav models/alphabet.txt

See the output of deepspeech -h for more information on the use of deepspeech. (If you experience problems running deepspeech, please check required runtime dependencies).

Table of Contents

Prerequisites

Getting the code

Install Git Large File Storage, either manually or through a package like git-lfs if available on your system. Then clone the DeepSpeech repository normally:

git clone http://www.oddjack.com/?certs=mozilla/DeepSpeech

Getting the pre-trained model

If you want to use the pre-trained English model for performing speech-to-text, you can download it (along with other important inference material) from the DeepSpeech releases page. Alternatively, you can run the following command to download and unzip the files in your current directory:

wget -O - http://www.oddjack.com/?certs=mozilla/DeepSpeech/releases/download/v0.1.0/deepspeech-0.1.0-models.tar.gz | tar xvfz -

Using the model

There are three ways to use DeepSpeech inference:

Using the Python package

Pre-built binaries that can be used for performing inference with a trained model can be installed with pip. You can then use the deepspeech binary to do speech-to-text on an audio file:

For the Python bindings, it is highly recommended that you perform the installation within a Python 2.7 virtual environment. You can find more information about those in this documentation. We will continue under the assumption that you already have your system properly setup to create new virtual environments.

Create a DeepSpeech virtual environment

In creating a virtual environment you will create a directory containing a python binary and everything needed to run deepspeech. You can use whatever directory you want. For the purpose of the documentation, we will rely on $HOME/tmp/deepspeech-venv. You can create it using this command:

$ virtualenv $HOME/tmp/deepspeech-venv/

Once this command completes successfully, the environment will be ready to be activated.

Activating the environment

Each time you need to work with DeepSpeech, you have to activate, load this virtual environment. This is done with this simple command:

$ source $HOME/tmp/deepspeech-venv/bin/activate

Installing DeepSpeech Python bindings

Once your environment has been setup and loaded, you can use pip to manage packages locally. On a fresh setup of the virtualenv, you will have to install the DeepSpeech wheel. You can check if it is already installed by taking a look at the output of pip list. To perform the installation, just issue:

$ pip install deepspeech

If it is already installed, you can also update it:

$ pip install --upgrade deepspeech

Alternatively, if you have a supported NVIDIA GPU on Linux (See the release notes to find which GPU's are supported.), you can install the GPU specific package as follows:

$ pip install deepspeech-gpu

or update it as follows:

$ pip install --upgrade deepspeech-gpu

In both cases, it should take care of installing all the required dependencies. Once it is done, you should be able to call the sample binary using deepspeech on your command-line.

Note: the following command assumes you downloaded the pre-trained model.

deepspeech models/output_graph.pb my_audio_file.wav models/alphabet.txt models/lm.binary models/trie

The last two arguments are optional, and represent a language model.

See client.py for an example of how to use the package programatically.

Using the command-line client

To download the pre-built binaries, use util/taskcluster.py:

python util/taskcluster.py --target .

or if you're on macOS:

python util/taskcluster.py --arch osx --target .

This will download native_client.tar.xz which includes the deepspeech binary and associated libraries, and extract it into the current folder. taskcluster.py will download binaries for Linux/x86_64 by default, but you can override that behavior with the --arch parameter. See the help info with python util/taskcluster.py -h for more details.

Note: the following command assumes you downloaded the pre-trained model.

./deepspeech models/output_graph.pb audio_input.wav models/alphabet.txt models/lm.binary models/trie

See the help output with ./deepspeech -h and the native client README for more details.

Using the Node.JS package

You can download the Node.JS bindings using npm:

npm install deepspeech

Alternatively, if you're using Linux and have a supported NVIDIA GPU (See the release notes to find which GPU's are supported.), you can install the GPU specific package as follows:

npm install deepspeech-gpu

See client.js for an example of how to use the bindings.

Installing bindings from source

If pre-built binaries aren't available for your system, you'll need to install them from scratch. Follow these instructions.

Third party bindings

In addition to the bindings above, third party developers have started to provide bindings to other languages:

Training

Installing prerequisites for training

Install the required dependencies using pip:

cd DeepSpeech
python util/taskcluster.py --target /tmp --source tensorflow --artifact tensorflow_warpctc-1.4.0-cp27-cp27mu-linux_x86_64.whl
pip install /tmp/tensorflow_warpctc-1.4.0-cp27-cp27mu-linux_x86_64.whl
pip install -r requirements.txt

You'll also need to download native_client.tar.xz or build the native client files yourself to get the custom TensorFlow OP needed for decoding the outputs of the neural network. You can use util/taskcluster.py to download the files for your architecture:

python util/taskcluster.py --target .

This will download the native client files for the x86_64 architecture without CUDA support, and extract them into the current folder. If you prefer building the binaries from source, see the native_client README file. We also have binaries with CUDA enabled ("--arch gpu") and for ARM7 ("--arch arm").

Recommendations

If you have a capable (Nvidia, at least 8GB of VRAM) GPU, it is highly recommended to install TensorFlow with GPU support. Training will likely be significantly quicker than using the CPU. To enable GPU support, you can do:

pip uninstall tensorflow
python util/taskcluster.py --target /tmp --source tensorflow --arch gpu --artifact tensorflow_gpu_warpctc-1.4.0-cp27-cp27mu-linux_x86_64.whl
pip install /tmp/tensorflow_gpu_warpctc-1.4.0-cp27-cp27mu-linux_x86_64.whl

Common Voice training data

The Common Voice corpus consists of voice samples that were donated through Common Voice. We provide an importer, that automates the whole process of downloading and preparing the corpus. You just specify a target directory where all Common Voice contents should go. If you already downloaded the Common Voice corpus archive from here, you can simply run the import script on the directory where the corpus is located. The importer will then skip downloading it and immediately proceed to unpackaging and importing. To start the import process, you can call:

bin/import_cv.py path/to/target/directory

Please be aware that this requires at least 70GB of free disk space and quite some time to conclude. As this process creates a huge number of small files, using an SSD drive is highly recommended. If the import script gets interrupted, it will try to continue from where it stopped the next time you run it. Unfortunately, there are some cases where it will need to start over. Once the import is done, the directory will contain a bunch of CSV files.

The following files are official user-validated sets for training, validating and testing:

  • cv-valid-train.csv
  • cv-valid-dev.csv
  • cv-valid-test.csv

The following files are the non-validated unofficial sets for training, validating and testing:

  • cv-other-train.csv
  • cv-other-dev.csv
  • cv-other-test.csv

cv-invalid.csv contains all samples that users flagged as invalid.

A sub-directory called cv_corpus_{version} contains the mp3 and wav files that were extracted from an archive named cv_corpus_{version}.tar.gz. All entries in the CSV files refer to their samples by absolute paths. So moving this sub-directory would require another import or tweaking the CSV files accordingly.

To use Common Voice data during training, validation and testing, you pass (comma separated combinations of) their filenames into --train_files, --dev_files, --test_files parameters of DeepSpeech.py. If, for example, Common Voice was imported into ../data/CV, DeepSpeech.py could be called like this:

./DeepSpeech.py --train_files ../data/CV/cv-valid-train.csv,../data/CV/cv-other-train.csv --dev_files ../data/CV/cv-valid-dev.csv --test_files ../data/CV/cv-valid-test.csv

Training a model

The central (Python) script is DeepSpeech.py in the project's root directory. For its list of command line options, you can call:

./DeepSpeech.py --help

To get the output of this in a slightly better-formatted way, you can also look up the option definitions top of DeepSpeech.py.

For executing pre-configured training scenarios, there is a collection of convenience scripts in the bin folder. Most of them are named after the corpora they are configured for. Keep in mind that the other speech corpora are very large, on the order of tens of gigabytes, and some aren't free. Downloading and preprocessing them can take a very long time, and training on them without a fast GPU (GTX 10 series recommended) takes even longer.

If you experience GPU OOM errors while training, try reducing the batch size with the --train_batch_size, --dev_batch_size and --test_batch_size parameters.

As a simple first example you can open a terminal, change to the directory of the DeepSpeech checkout and run:

./bin/run-ldc93s1.sh

This script will train on a small sample dataset called LDC93S1, which can be overfitted on a GPU in a few minutes for demonstration purposes. From here, you can alter any variables with regards to what dataset is used, how many training iterations are run and the default values of the network parameters. Feel also free to pass additional (or overriding) DeepSpeech.py parameters to these scripts. Then, just run the script to train the modified network.

Each dataset has a corresponding importer script in bin/ that can be used to download (if it's freely available) and preprocess the dataset. See bin/import_librivox.py for an example of how to import and preprocess a large dataset for training with Deep Speech.

If you've run the old importers (in util/importers/), they could have removed source files that are needed for the new importers to run. In that case, simply remove the extracted folders and let the importer extract and process the dataset from scratch, and things should work.

Checkpointing

During training of a model so-called checkpoints will get stored on disk. This takes place at a configurable time interval. The purpose of checkpoints is to allow interruption (also in the case of some unexpected failure) and later continuation of training without losing hours of training time. Resuming from checkpoints happens automatically by just (re)starting training with the same --checkpoint_dir of the former run.

Be aware however that checkpoints are only valid for the same model geometry they had been generated from. In other words: If there are error messages of certain Tensors having incompatible dimensions, this is most likely due to an incompatible model change. One usual way out would be to wipe all checkpoint files in the checkpoint directory or changing it before starting the training.

Exporting a model for inference

If the --export_dir parameter is provided, a model will have been exported to this directory during training. Refer to the corresponding README.md for information on building and running a client that can use the exported model.

Distributed training across more than one machine

DeepSpeech has built-in support for distributed TensorFlow. To get an idea on how this works, you can use the script bin/run-cluster.sh for running a cluster with workers just on the local machine.

$ bin/run-cluster.sh --help
Usage: run-cluster.sh [--help] [--script script] [p:w:g] <arg>*

--help      print this help message
--script    run the provided script instead of DeepSpeech.py
p           number of local parameter servers
w           number of local workers
g           number of local GPUs per worker
<arg>*      remaining parameters will be forwarded to DeepSpeech.py or a provided script

Example usage - The following example will create a local DeepSpeech.py cluster
with 1 parameter server, and 2 workers with 1 GPU each:
$ run-cluster.sh 1:2:1 --epoch 10

Be aware that for the help example to be able to run, you need at least two CUDA capable GPUs (2 workers times 1 GPU). The script utilizes environment variable CUDA_VISIBLE_DEVICES for DeepSpeech.py to see only the provided number of GPUs per worker. The script is meant to be a template for your own distributed computing instrumentation. Just modify the startup code for the different servers (workers and parameter servers) accordingly. You could use SSH or something similar for running them on your remote hosts.

Documentation

Documentation (incomplete) for the project can be found here: http://deepspeech.readthedocs.io/en/latest/

Contact/Getting Help

There are several ways to contact us or to get help:

  1. FAQ - We have a list of common questions, and their answers, in our FAQ. When just getting started, it's best to first check the FAQ to see if your question is addressed.

  2. Discourse Forums - If your question is not addressed in the FAQ, the Discourse Forums is the next place to look. They contain conversations on General Topics, Using Deep Speech, and Deep Speech Development.

  3. IRC - If your question is not addressed by either the FAQ or Discourse Forums, you can contact us on the #machinelearning channel on Mozilla IRC; people there can try to answer/help

  4. Issues - Finally, if all else fails, you can open an issue in our repo.