Automatic speech recognition (ASR) system implementation that utilizes the connectionist temporal classification (CTC) cost function. It's inspired by Baidu's Deep Speech: Scaling up end-to-end speech recognition and Deep Speech 2: End-to-End Speech Recognition in English and Mandarin papers. The system is trained on a combined corpus, containing 900+ hours. It achieves a word error rate (WER) of 12.6% on the test dataset, without the use of an external language model.
(a) shows the Deep Speech (1) model and (b) a version of the Deep Speech 2 model architecture.
The system was tested on Arch Linux and Ubuntu 16.04, with Python version 3.5+ and the 1.12.0 version of TensorFlow. It's highly recommended to use TensorFlow with GPU support for training.
# Install dependencies.
sudo pacman -S sox python-tensorflow-opt-cuda tensorbaord
# Install optional dependencies. LaTeX is only required to plot nice looking graphs.
sudo pacman -S texlive-most
# Clone reposetory and install Python depdendencies.
git clone https://github.com/mdangschat/ctc-asr.git
cd speech
git checkout <release_tag>
# Setup optional virtual environment.
pip install -r requirements.txt
Be aware that the requirements.txt file lists tensorflow as dependency,
if you install TensorFlow through pip consider removing it as
dependency and install tensorflow-gpu instead.
It could also be worth it to build TensorFlow from source.
# Install dependencies.
sudo apt install python3-tk sox libsox-fmt-all
# Install optional dependencies. LaTeX is only required to plot nice looking graphs.
sudo apt install texlive
# Clone reposetory and install Python depdendencies. Don't forget to use tensorflow-gpu.
git clone https://github.com/mdangschat/ctc-asr.git
cd speech
git checkout <release_tag>
# Setup optional virtual environment.
pip3 install -r requirements.txt
The network architecture and training parameters can be configured by adding the appropriate flags
or by directly editing the asr/params.py configuration file.
The default configuration requires quite a lot of VRAM (about 16 GB), consider reducing the number of units per
layer (num_units_dense, num_units_rnn) and the amount of RNN layers (num_layers_rnn).
There is list of some free speech corpora at the end of this section. However, the corpus is not part of this repository and has to be acquired by each user. For a quick start there is the speech-corpus-dl helper, that downloads a few free corpora, prepares the data and creates a merged corpus.
All audio files have to be 16 kHz, mono, WAV files. For my trainings, I removed examples shorter than 0.7 and longer than 17.0 seconds. Additionally, TEDLIUM examples with labels of fewer than 5 words have also been removed.
The following tree shows a possible structure for the required directories:
./ctc-asr
├── asr
├── [...]
├── LICENSE
├── README.md
├── requirements.txt
├── testruns.md
./ctc-asr-checkpoints
└── 3c2r2d-rnn
├── [...]
./speech-corpus
├── cache
├── corpus
│ ├── cvv2
│ ├── LibriSpeech
│ ├── tatoeba_audio_eng
│ └── TEDLIUM_release2
├── corpus.json
├── dev.csv
├── test.csv
└── train.csv
Assuming that this repository is cloned into some/folder/ctc-asr, then by default
the CSV files are expected to be in some/folder/speech-corpus and the audio files in
some/folder/speech-corpus/corpus.
TensorFlow checkpoints are written into some/folder/ctc-asr-checkpoints.
Both folders (ctc-asr-checkpoints and speech-corpus) must exist, they can be changed
in the asr/params.py file.
The CSV files (e.g. train.csv) have the following format:
path;label;length
relative/path/to/example;lower case transcription without puntuation;3.14159265359
[...]
Where path is the relative WAV path from the DATA_DIR/corpus/ directory (String).
By default, label is the lower case transcription without punctuation (String).
Finally, length is the audio length in seconds (Float).
- Common Voice (v1)
- LibriSpeech ASR Corpus
- Tatoeba
- TED-Lium (v2)
- TIMIT
ipython python/dataset/word_counts.py
Calculating statistics for /home/gpuinstall/workspace/ctc-asr/data/train.csv
Word based statistics:
total_words = 10,069,671
number_unique_words = 81,161
mean_sentence_length = 14.52 words
min_sentence_length = 1 words
max_sentence_length = 84 words
Most common words: [('the', 551055), ('to', 306197), ('and', 272729), ('of', 243032), ('a', 223722), ('i', 192151), ('in', 149797), ('that', 146820), ('you', 144244), ('it', 118133)]
27416 words occurred only 1 time; 37,422 words occurred only 2 times; 49,939 words occurred only 5 times; 58,248 words occurred only 10 times.
Character based statistics:
total_characters = 52,004,043
mean_label_length = 75.00 characters
min_label_length = 2 characters
max_label_length = 422 characters
Most common characters: [(' ', 9376326), ('e', 5264177), ('t', 4205041), ('o', 3451023), ('a', 3358945), ('i', 2944773), ('n', 2858788), ('s', 2624239), ('h', 2598897), ('r', 2316473), ('d', 1791668), ('l', 1686896), ('u', 1234080), ('m', 1176076), ('w', 1052166), ('c', 999590), ('y', 974918), ('g', 888446), ('f', 851710), ('p', 710252), ('b', 646150), ('v', 421126), ('k', 387714), ('x', 62547), ('j', 61048), ('q', 34558), ('z', 26416)]
Most common characters: [' ', 'e', 't', 'o', 'a', 'i', 'n', 's', 'h', 'r', 'd', 'l', 'u', 'm', 'w', 'c', 'y', 'g', 'f', 'p', 'b', 'v', 'k', 'x', 'j', 'q', 'z']
Start training by invoking asr/train.py.
Use asr/train.py -- --delete to start a clean run and remove the old checkpoints.
Please note that all commands are expected to be executed from the projects root folder.
The additional -- before the actual flags begin is used to indicate the end of IPython flags.
The training progress can be monitored using Tensorboard.
To start Tensorboard use tensorboard --logdir <checkpoint_directory>.
By default it can then be accessed via localhost:6006.
Evaluate the current model by invoking asr/evaluate.py.
Use asr/evaluate.py -- --dev to run on the development dataset, instead of the test set.
To evaluate a given 16 kHz, mono WAV file use asr/predict.py --input <wav_path>.