OpenSpeech provides reference implementations of various ASR modeling papers and three languages recipe to perform tasks on automatic speech recognition. We aim to make ASR technology easier to use for everyone.
OpenSpeech is backed by the two powerful libraries — PyTorch-Lightning and Hydra.
Various features are available in the above two libraries, including Multi-GPU and TPU training, Mixed-precision, and hierarchical configuration management.
We appreciate any kind of feedback or contribution. Feel free to proceed with small issues like bug fixes, documentation improvement. For major contributions and new features, please discuss with the collaborators in corresponding issues.
- May 2022 openspeech 0.4.0 released
- Aug 2021 Added Smart Batching
- Jul 2021 openspeech 0.3.0 released
- Jul 2021 Added transducer beam search logic
- Jun 2021 Added ContextNet
- Jun 2021 Added language model training pipeline
- Jun 2021 openspeech 0.1.0 released
OpenSpeech is a framework for making end-to-end speech recognizers. End-to-end (E2E) automatic speech recognition (ASR) is an emerging paradigm in the field of neural network-based speech recognition that offers multiple benefits. Traditional “hybrid” ASR systems, which are comprised of an acoustic model, language model, and pronunciation model, require separate training of these components, each of which can be complex.
For example, training of an acoustic model is a multi-stage process of model training and time alignment between the speech acoustic feature sequence and output label sequence. In contrast, E2E ASR is a single integrated approach with a much simpler training pipeline with models that operate at low audio frame rates. This reduces the training time, decoding time, and allows joint optimization with downstream processing such as natural language understanding.
Because of these advantages, many end-to-end speech recognition related open sources have emerged. But, Many of them are based on basic PyTorch or Tensorflow, it is very difficult to use various functions such as mixed-precision, multi-node training, and TPU training etc. However, with frameworks such as PyTorch-Lighting, these features can be easily used. So we have created a speech recognition framework that introduced PyTorch-Lightning and Hydra for easy use of these advanced features.
- PyTorch-Lighting base framework.
- Various functions: mixed-precision, multi-node training, tpu training etc.
- Models become hardware agnostic
- Make fewer mistakes because lightning handles the tricky engineering
- Lightning has dozens of integrations with popular machine learning tools.
- Easy-to-experiment with the famous ASR models.
- Supports 20+ models and is continuously updated.
- Low barrier to entry for educators and practitioners.
- Save time for researchers who want to conduct various experiments.
- Provides recipes for the most widely used languages, English, Chinese, and + Korean.
- LibriSpeech - 1,000 hours of English dataset most widely used in ASR tasks.
- AISHELL-1 - 170 hours of Chinese Mandarin speech corpus.
- KsponSpeech - 1,000 hours of Korean open-domain dialogue speech.
- Easily customize a model or a new dataset to your needs:
- The default hparams of the supported models are provided but can be easily adjusted.
- Easily create a custom model by combining modules that are already provided.
- If you want to use the new dataset, you only need to define a
pl.LightingDataModule
andTokenizer
classes.
- Audio processing
- Representative audio features such as Spectrogram, Mel-Spectrogram, Filter-Bank, and MFCC can be used easily.
- Provides a variety of augmentation, including SpecAugment, Noise Injection, and Audio Joining.
- This framework provides code for training ASR models, but does not provide APIs by pre-trained models.
- This framework does not provides pre-trained models.
We support all the models below. Note that, the important concepts of the model have been implemented to match, but the details of the implementation may vary.
- DeepSpeech2 (from Baidu Research) released with paper Deep Speech 2: End-to-End Speech Recognition in English and Mandarin, by Dario Amodei, Rishita Anubhai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Jingdong Chen, Mike Chrzanowski, Adam Coates, Greg Diamos, Erich Elsen, Jesse Engel, Linxi Fan, Christopher Fougner, Tony Han, Awni Hannun, Billy Jun, Patrick LeGresley, Libby Lin, Sharan Narang, Andrew Ng, Sherjil Ozair, Ryan Prenger, Jonathan Raiman, Sanjeev Satheesh, David Seetapun, Shubho Sengupta, Yi Wang, Zhiqian Wang, Chong Wang, Bo Xiao, Dani Yogatama, Jun Zhan, Zhenyao Zhu.
- RNN-Transducer (from University of Toronto) released with paper Sequence Transduction with Recurrent Neural Networks, by Alex Graves.
- LSTM Language Model (from RWTH Aachen University) released with paper LSTM Neural Networks for Language Modeling, by Martin Sundermeyer, Ralf Schluter, and Hermann Ney.
- Listen Attend Spell (from Carnegie Mellon University and Google Brain) released with paper Listen, Attend and Spell, by William Chan, Navdeep Jaitly, Quoc V. Le, Oriol Vinyals.
- Location-aware attention based Listen Attend Spell (from University of Wrocław and Jacobs University and Universite de Montreal) released with paper Attention-Based Models for Speech Recognition, by Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, Yoshua Bengio.
- Joint CTC-Attention based Listen Attend Spell (from Mitsubishi Electric Research Laboratories and Carnegie Mellon University) released with paper Joint CTC-Attention based End-to-End Speech Recognition using Multi-task Learning, by Suyoun Kim, Takaaki Hori, Shinji Watanabe.
- Deep CNN Encoder with Joint CTC-Attention Listen Attend Spell (from Mitsubishi Electric Research Laboratories and Massachusetts Institute of Technology and Carnegie Mellon University) released with paper Advances in Joint CTC-Attention based End-to-End Speech Recognition with a Deep CNN Encoder and RNN-LM, by Takaaki Hori, Shinji Watanabe, Yu Zhang, William Chan.
- Multi-head attention based Listen Attend Spell (from Google) released with paper State-of-the-art Speech Recognition With Sequence-to-Sequence Models, by Chung-Cheng Chiu, Tara N. Sainath, Yonghui Wu, Rohit Prabhavalkar, Patrick Nguyen, Zhifeng Chen, Anjuli Kannan, Ron J. Weiss, Kanishka Rao, Ekaterina Gonina, Navdeep Jaitly, Bo Li, Jan Chorowski, Michiel Bacchiani.
- Speech-Transformer (from University of Chinese Academy of Sciences and Institute of Automation and Chinese Academy of Sciences) released with paper Speech-Transformer: A No-Recurrence Sequence-to-Sequence Model for Speech Recognition, by Linhao Dong; Shuang Xu; Bo Xu.
- VGG-Transformer (from Facebook AI Research) released with paper Transformers with convolutional context for ASR, by Abdelrahman Mohamed, Dmytro Okhonko, Luke Zettlemoyer.
- Transformer with CTC (from NTT Communication Science Laboratories, Waseda University, Center for Language and Speech Processing, Johns Hopkins University) released with paper Improving Transformer-based End-to-End Speech Recognition with Connectionist Temporal Classification and Language Model Integration, by Shigeki Karita, Nelson Enrique Yalta Soplin, Shinji Watanabe, Marc Delcroix, Atsunori Ogawa, Tomohiro Nakatani.
- Joint CTC-Attention based Transformer (from NTT Corporation) released with paper Self-Distillation for Improving CTC-Transformer-based ASR Systems, by Takafumi Moriya, Tsubasa Ochiai, Shigeki Karita, Hiroshi Sato, Tomohiro Tanaka, Takanori Ashihara, Ryo Masumura, Yusuke Shinohara, Marc Delcroix.
- Transformer Language Model (from Amazon Web Services) released with paper Language Models with Transformers, by Chenguang Wang, Mu Li, Alexander J. Smola.
- Jasper (from NVIDIA and New York University) released with paper Jasper: An End-to-End Convolutional Neural Acoustic Model, by Jason Li, Vitaly Lavrukhin, Boris Ginsburg, Ryan Leary, Oleksii Kuchaiev, Jonathan M. Cohen, Huyen Nguyen, Ravi Teja Gadde.
- QuartzNet (from NVIDIA and Univ. of Illinois and Univ. of Saint Petersburg) released with paper QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions, by Samuel Kriman, Stanislav Beliaev, Boris Ginsburg, Jocelyn Huang, Oleksii Kuchaiev, Vitaly Lavrukhin, Ryan Leary, Jason Li, Yang Zhang.
- Transformer Transducer (from Facebook AI) released with paper Transformer-Transducer: End-to-End Speech Recognition with Self-Attention, by Ching-Feng Yeh, Jay Mahadeokar, Kaustubh Kalgaonkar, Yongqiang Wang, Duc Le, Mahaveer Jain, Kjell Schubert, Christian Fuegen, Michael L. Seltzer.
- Conformer (from Google) released with paper Conformer: Convolution-augmented Transformer for Speech Recognition, by Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, Ruoming Pang.
- Conformer with CTC (from Northwestern Polytechnical University and University of Bordeaux and Johns Hopkins University and Human Dataware Lab and Kyoto University and NTT Corporation and Shanghai Jiao Tong University and Chinese Academy of Sciences) released with paper Recent Developments on ESPNET Toolkit Boosted by Conformer, by Pengcheng Guo, Florian Boyer, Xuankai Chang, Tomoki Hayashi, Yosuke Higuchi, Hirofumi Inaguma, Naoyuki Kamo, Chenda Li, Daniel Garcia-Romero, Jiatong Shi, Jing Shi, Shinji Watanabe, Kun Wei, Wangyou Zhang, Yuekai Zhang.
- Conformer with LSTM Decoder (from IBM Research AI) released with paper On the limit of English conversational speech recognition, by Zoltán Tüske, George Saon, Brian Kingsbury.
- ContextNet (from Google) released with paper ContextNet: Improving Convolutional Neural Networks for Automatic Speech Recognition with Global Context, by Wei Han, Zhengdong Zhang, Yu Zhang, Jiahui Yu, Chung-Cheng Chiu, James Qin, Anmol Gulati, Ruoming Pang, Yonghui Wu.
- Squeezeformer (from University of Berkeley) released with paper Squeezeformer: An Efficient Transformer for Automatic Speech Recognition, by Sehoon Kim, Amir Gholami, Albert Shaw, Nicholas Lee, Karttikeya Mangalam, Jitendra Malik, Michael W. Mahoney, Kurt Keutzer.
We use Hydra to control all the training configurations. If you are not familiar with Hydra we recommend visiting the Hydra website. Generally, Hydra is an open-source framework that simplifies the development of research applications by providing the ability to create a hierarchical configuration dynamically. If you want to know how we used Hydra, we recommend you to read here.
We support LibriSpeech, KsponSpeech, and AISHELL-1.
LibriSpeech is a corpus of approximately 1,000 hours of 16kHz read English speech, prepared by Vassil Panayotov with the assistance of Daniel Povey. The data was derived from reading audiobooks from the LibriVox project, and has been carefully segmented and aligned.
Aishell is an open-source Chinese Mandarin speech corpus published by Beijing Shell Shell Technology Co.,Ltd. 400 people from different accent areas in China were invited to participate in the recording, which was conducted in a quiet indoor environment using high fidelity microphone and downsampled to 16kHz.
KsponSpeech is a large-scale spontaneous speech corpus of Korean. This corpus contains 969 hours of general open-domain dialog utterances, spoken by about 2,000 native Korean speakers in a clean environment. All data were constructed by recording the dialogue of two people freely conversing on a variety of topics and manually transcribing the utterances. To start training, the KsponSpeech dataset must be prepared in advance. To download KsponSpeech, you need permission from AI Hub.
- Acoustic model manifest file format:
LibriSpeech/test-other/8188/269288/8188-269288-0052.flac ▁ANNIE ' S ▁MANNER ▁WAS ▁VERY ▁MYSTERIOUS 4039 20 5 531 17 84 2352
LibriSpeech/test-other/8188/269288/8188-269288-0053.flac ▁ANNIE ▁DID ▁NOT ▁MEAN ▁TO ▁CONFIDE ▁IN ▁ANYONE ▁THAT ▁NIGHT ▁AND ▁THE ▁KIND EST ▁THING ▁WAS ▁TO ▁LEAVE ▁HER ▁A LONE 4039 99 35 251 9 4758 11 2454 16 199 6 4 323 200 255 17 9 370 30 10 492
LibriSpeech/test-other/8188/269288/8188-269288-0054.flac ▁TIRED ▁OUT ▁LESLIE ▁HER SELF ▁DROPP ED ▁A SLEEP 1493 70 4708 30 115 1231 7 10 1706
LibriSpeech/test-other/8188/269288/8188-269288-0055.flac ▁ANNIE ▁IS ▁THAT ▁YOU ▁SHE ▁CALL ED ▁OUT 4039 34 16 25 37 208 7 70
LibriSpeech/test-other/8188/269288/8188-269288-0056.flac ▁THERE ▁WAS ▁NO ▁REPLY ▁BUT ▁THE ▁SOUND ▁OF ▁HURRY ING ▁STEPS ▁CAME ▁QUICK ER ▁AND ▁QUICK ER ▁NOW ▁AND ▁THEN ▁THEY ▁WERE ▁INTERRUPTED ▁BY ▁A ▁GROAN 57 17 56 1368 33 4 489 8 1783 14 1381 133 571 49 6 571 49 82 6 76 45 54 2351 44 10 3154
LibriSpeech/test-other/8188/269288/8188-269288-0057.flac ▁OH ▁THIS ▁WILL ▁KILL ▁ME ▁MY ▁HEART ▁WILL ▁BREAK ▁THIS ▁WILL ▁KILL ▁ME 299 46 71 669 50 41 235 71 977 46 71 669 50
...
...
You can simply train with LibriSpeech dataset like below:
- Example1: Train the
conformer-lstm
model withfilter-bank
features on GPU.
$ python3 ./openspeech_cli/hydra_train.py \
dataset=librispeech \
dataset.dataset_download=True \
dataset.dataset_path=$DATASET_PATH \
dataset.manifest_file_path=$MANIFEST_FILE_PATH \
tokenizer=libri_subword \
model=conformer_lstm \
audio=fbank \
lr_scheduler=warmup_reduce_lr_on_plateau \
trainer=gpu \
criterion=cross_entropy
You can simply train with KsponSpeech dataset like below:
- Example2: Train the
listen-attend-spell
model withmel-spectrogram
features On TPU:
$ python3 ./openspeech_cli/hydra_train.py \
dataset=ksponspeech \
dataset.dataset_path=$DATASET_PATH \
dataset.manifest_file_path=$MANIFEST_FILE_PATH \
dataset.test_dataset_path=$TEST_DATASET_PATH \
dataset.test_manifest_dir=$TEST_MANIFEST_DIR \
tokenizer=kspon_character \
model=listen_attend_spell \
audio=melspectrogram \
lr_scheduler=warmup_reduce_lr_on_plateau \
trainer=tpu \
criterion=cross_entropy
You can simply train with AISHELL-1 dataset like below:
- Example3: Train the
quartznet
model withmfcc
features On GPU with FP16:
$ python3 ./openspeech_cli/hydra_train.py \
dataset=aishell \
dataset.dataset_path=$DATASET_PATH \
dataset.dataset_download=True \
dataset.manifest_file_path=$MANIFEST_FILE_PATH \
tokenizer=aishell_character \
model=quartznet15x5 \
audio=mfcc \
lr_scheduler=warmup_reduce_lr_on_plateau \
trainer=gpu-fp16 \
criterion=ctc
- Example1: Evaluation the
listen_attend_spell
model:
$ python3 ./openspeech_cli/hydra_eval.py \
audio=melspectrogram \
eval.dataset_path=$DATASET_PATH \
eval.checkpoint_path=$CHECKPOINT_PATH \
eval.manifest_file_path=$MANIFEST_FILE_PATH \
model=listen_attend_spell \
tokenizer=kspon_character \
tokenizer.vocab_path=$VOCAB_FILE_PATH \
- Example2: Evaluation the
listen_attend_spell
,conformer_lstm
models with ensemble:
$ python3 ./openspeech_cli/hydra_eval.py \
audio=melspectrogram \
eval.model_names=(listen_attend_spell, conformer_lstm) \
eval.dataset_path=$DATASET_PATH \
eval.checkpoint_paths=($CHECKPOINT_PATH1, $CHECKPOINT_PATH2) \
eval.ensemble_weights=(0.3, 0.7) \
eval.ensemble_method=weighted \
eval.manifest_file_path=$MANIFEST_FILE_PATH
dataset.dataset_path
: $BASE_PATH/KsponSpeech
$BASE_PATH/KsponSpeech
├── KsponSpeech_01
├── KsponSpeech_02
├── KsponSpeech_03
├── KsponSpeech_04
└── KsponSpeech_05
dataset.test_dataset_path
: $BASE_PATH/KsponSpeech_eval
$BASE_PATH/KsponSpeech_eval
├── eval_clean
└── eval_other
dataset.test_manifest_dir
: $BASE_PATH/KsponSpeech_scripts
$BASE_PATH/KsponSpeech_scripts
├── eval_clean.trn
└── eval_other.trn
Language model training requires only data to be prepared in the following format:
openspeech is a framework for making end-to-end speech recognizers.
end to end automatic speech recognition is an emerging paradigm in the field of neural network-based speech recognition that offers multiple benefits.
because of these advantages, many end-to-end speech recognition related open sources have emerged.
...
...
Note that you need to use the same vocabulary as the acoustic model.
- Example: Train the
lstm_lm
model:
$ python3 ./openspeech_cli/hydra_lm_train.py \
dataset=lm \
dataset.dataset_path=../../../lm.txt \
tokenizer=kspon_character \
tokenizer.vocab_path=../../../labels.csv \
model=lstm_lm \
lr_scheduler=tri_stage \
trainer=gpu \
criterion=perplexity
This project recommends Python 3.7 or higher. We recommend creating a new virtual environment for this project (using virtual env or conda).
- numpy:
pip install numpy
(Refer here for problem installing Numpy). - pytorch: Refer to PyTorch website to install the version w.r.t. your environment.
- librosa:
conda install -c conda-forge librosa
(Refer here for problem installing librosa) - torchaudio:
pip install torchaudio==0.6.0
(Refer here for problem installing torchaudio) - sentencepiece:
pip install sentencepiece
(Refer here for problem installing sentencepiece) - pytorch-lightning:
pip install pytorch-lightning
(Refer here for problem installing pytorch-lightning) - hydra:
pip install hydra-core --upgrade
(Refer here for problem installing hydra) - warp-rnnt: Refer to warp-rnnt page to install the library.
- ctcdecode: Refer to ctcdecode page to install the library.
You can install OpenSpeech with pypi.
pip install openspeech-core
Currently we only support installation from source code using setuptools. Checkout the source code and run the following commands:
$ ./install.sh
For faster training install NVIDIA's apex library:
$ git clone https://github.com/NVIDIA/apex
$ cd apex
# ------------------------
# OPTIONAL: on your cluster you might need to load CUDA 10 or 9
# depending on how you installed PyTorch
# see available modules
module avail
# load correct CUDA before install
module load cuda-10.0
# ------------------------
# make sure you've loaded a cuda version > 4.0 and < 7.0
module load gcc-6.1.0
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
If you have any questions, bug reports, and feature requests, please open an issue on Github.
We appreciate any kind of feedback or contribution. Feel free to proceed with small issues like bug fixes, documentation improvement. For major contributions and new features, please discuss with the collaborators in corresponding issues.
We follow PEP-8 for code style. Especially the style of docstrings is important to generate documentation.
This project is licensed under the MIT LICENSE - see the LICENSE.md file for details
If you use the system for academic work, please cite:
@GITHUB{2021-OpenSpeech,
author = {Kim, Soohwan and Ha, Sangchun and Cho, Soyoung},
author email = {sh951011@gmail.com, seomk9896@gmail.com, soyoung.cho@kaist.ac.kr}
title = {OpenSpeech: Open-Source Toolkit for End-to-End Speech Recognition},
howpublished = {\url{https://github.com/openspeech-team/openspeech}},
docs = {\url{https://openspeech-team.github.io/openspeech}},
year = {2021}
}