Espresso is an open-source, modular, extensible end-to-end neural automatic speech recognition (ASR) toolkit based on the deep learning library PyTorch and the popular neural machine translation toolkit fairseq. Espresso supports distributed training across GPUs and computing nodes, and features various decoding approaches commonly employed in ASR, including look-ahead word-based language model fusion, for which a fast, parallelized decoder is implemented.
We provide state-of-the-art training recipes for the following speech datasets:
- September 2022: CTC model training and decoding are supported. Check out a config file example.
- February 2022: Conformer encoder is implemented. Simply add one line option in the config file to enable it. See examples: here and here.
- December 2021: A suite of Transducer model training and decoding code is added. An illustrative LibriSpeech recipe is here. The training requires torchaudio >= 0.10.0 installed.
- April 2021: On-the-fly feature extraction from raw waveforms with torchaudio is supported. A LibriSpeech recipe is released here with no dependency on Kaldi and using YAML files (via Hydra) for configuring experiments.
- June 2020: Transformer recipes released.
- April 2020: Both E2E LF-MMI (using PyChain) and Cross-Entropy training for hybrid ASR are now supported. WSJ recipes are provided here and here as examples, respectively.
- March 2020: SpecAugment is supported and relevant recipes are released.
- September 2019: We are in an effort of isolating Espresso from fairseq, resulting in a standalone package that can be directly
pip installed.
- PyTorch version >= 1.10.0
- Python version >= 3.8
- For training new models, you'll also need an NVIDIA GPU and NCCL
- To install Espresso from source and develop locally:
git clone https://github.com/freewym/espresso
cd espresso
pip install --editable .
# on MacOS:
# CFLAGS="-stdlib=libc++" pip install --editable ./
pip install kaldi_io sentencepiece soundfile
cd espresso/tools; make KALDI=<path/to/a/compiled/kaldi/directory>add your Python path to PATH variable in examples/asr_<dataset>/path.sh, the current default is ~/anaconda3/bin.
kaldi_io is required for reading kaldi scp files. sentencepiece is required for subword pieces training/encoding. soundfile is required for reading raw waveform files. Kaldi is required for data preparation, feature extraction, scoring for some datasets (e.g., Switchboard), and decoding for all hybrid systems.
edit PYTHON_DIR variable in espresso/tools/Makefile (default: ~/anaconda3/bin), and then
cd espresso/tools; make openfst pychain- For faster training install NVIDIA's apex library:
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
--global-option="--deprecated_fused_adam" --global-option="--xentropy" \
--global-option="--fast_multihead_attn" ./Espresso is MIT-licensed.
Please cite Espresso as:
@inproceedings{wang2019espresso,
title = {Espresso: A Fast End-to-end Neural Speech Recognition Toolkit},
author = {Yiming Wang and Tongfei Chen and Hainan Xu
and Shuoyang Ding and Hang Lv and Yiwen Shao
and Nanyun Peng and Lei Xie and Shinji Watanabe
and Sanjeev Khudanpur},
booktitle = {2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
year = {2019},
}