End-to-end speech recognition using RNN-Transducer in Tensorflow 2.0

This speech recognition model is based off Google's Streaming End-to-end Speech Recognition For Mobile Devices research paper and is implemented in Python 3 using Tensorflow 2.0

To setup your environment, run the following command:

git clone --recurse https://github.com/noahchalifour/rnnt-speech-recognition.git
cd rnnt-speech-recognition
pip install tensorflow==2.2.0 # or tensorflow-gpu==2.2.0 for GPU support
pip install -r requirements.txt
./scripts/build_rnnt.sh # to setup the rnnt loss

You can find and download the Common Voice dataset here

Before you can train a model on the Common Voice dataset, you must first convert all the audio mp3 filetypes to wavs. Do so by running the following command:

NOTE: Make sure you have ffmpeg installed on your computer, as it uses that to convert mp3 to wav

./scripts/common_voice_convert.sh <data_dir> <# of threads>
python scripts/remove_missing_samples.py \
    --data_dir <data_dir> \
    --replace_old

After converting all the mp3s to wavs you need to preprocess the dataset, you can do so by running the following command:

python preprocess_common_voice.py \
    --data_dir <data_dir> \
    --output_dir <preprocessed_dir>

To train a simple model, run the following command:

python run_rnnt.py \
    --mode train \
    --data_dir <path to data directory>