Skip to content

Latest commit

 

History

History
139 lines (91 loc) · 5.77 KB

README.md

File metadata and controls

139 lines (91 loc) · 5.77 KB
Raw

PIKA: a lightweight speech processing toolkit based on Pytorch and (Py)Kaldi

PIKA is a lightweight speech processing toolkit based on Pytorch and (Py)Kaldi. The first release focuses on end-to-end speech recognition. We use Pytorch as deep learning engine, Kaldi for data formatting and feature extraction.

Key Features

  • On-the-fly data augmentation and feature extraction loader

  • TDNN Transformer encoder and convolution and transformer based decoder model structure

  • RNNT training and batch decoding

  • RNNT decoding with external Ngram FSTs (on-the-fly rescoring, aka, shallow fusion)

  • RNNT Minimum Bayes Risk (MBR) training

  • LAS forward and backward rescorer for RNNT

  • Efficient BMUF (Block model update filtering) based distributed training

Installation and Dependencies

In general, we recommend Anaconda since it comes with most dependencies. Other major dependencies include,

Pytorch

Please go to https://pytorch.org/ for pytorch installation, codes and scripts should be able to run against pytorch 0.4.0 and above. But we recommend 1.0.0 above for compatibility with RNNT loss module (see below)

Pykaldi and Kaldi

We use Kaldi (https://github.com/kaldi-asr/kaldi)) and PyKaldi (a python wrapper for Kaldi) for data processing, feature extraction and FST manipulations. Please go to Pykaldi website https://github.com/pykaldi/pykaldi for installation and make sure to build Pykaldi with ninja for efficiency. After following the installation process of pykaldi, you should have both Kaldi and Pykaldi dependencies ready.

CUDA-Warp RNN-Transducer

For RNNT loss module, we adopt the pytorch binding at https://github.com/1ytic/warp-rnnt

Others

Check requirements.txt for other dependencies.

Get Started

To get started, check all the training and decoding scripts located in egs directory.

I. Data preparation and RNNT training

egs/train_transducer_bmuf_otfaug.sh contains data preparation and RNNT training. One need to prepare training data and specify the training data directory,

#training data dir must contain wav.scp and label.txt files
#wav.scp: standard kaldi wav.scp file, see https://kaldi-asr.org/doc/data_prep.html 
#label.txt: label text file, the format is, uttid sequence-of-integer, where integer
#           is one-based indexing mapped label, note that zero is reserved for blank,  
#           ,eg., utt_id_1 3 5 7 10 23 
train_data_dir=

II. Continue with MBR training

With RNNT trained model, one can continued MBR training with egs/train_transducer_mbr_bmuf_otfaug.sh (assuming using the same training data, therefore data preparation is omitted). Make sure to specify the initial model,

--verbose \
--optim sgd \
--init_model $exp_dir/init.model \
--rnnt_scale 1.0 \
--sm_scale 0.8 \

III. Training LAS forward and backward rescorer

One can train a forward and backward LAS rescorer for your RNN-T model using egs/train_las_rescorer_bmuf_otfaug.sh. The LAS rescorer will share the encoder part with RNNT model, and has extra two-layer LSTM as additional encoder, make sure to specify the encoder sharing as,

--num_batches_per_epoch 526264 \
--shared_encoder_model $exp_dir/final.model \
--num_epochs 5 \

We support bi-directional LAS rescoring, i.e., forward and backward rescoring. Backward (right-to-left) rescoring is achieved by reversing sequential labels when conducting LAS model training. One can easily perform a backward LAS rescorer training by specifying,

--reverse_labels

IV. Decoding

egs/eval_transducer.sh is the main evluation script, which contains the decoding pipeline. Forward and backward LAS rescoring can be enabled by specifying these two models,

##########configs#############
#rnn transducer model
rnnt_model=
#forward and backward las rescorer model
lasrescorer_fw=
lasrescorer_bw=

Caveats

All the training and decoding hyper-parameters are adopted based on large-scale (e.g., 60khrs) training and internal evaluation data. One might need to re-tune hyper-parameters to acheive optimal performances. Also the WER (CER) scoring script is based on a Mandarin task, we recommend those who work on different languages rewrite scoring scripts.

References

[1] Improving Attention Based Sequence-to-Sequence Models for End-to-End English Conversational Speech Recognition, Chao Weng, Jia Cui, Guangsen Wang, Jun Wang, Chengzhu Yu, Dan Su, Dong Yu, InterSpeech 2018

[2] Minimum Bayes Risk Training of RNN-Transducer for End-to-End Speech Recognition, Chao Weng, Chengzhu Yu, Jia Cui, Chunlei Zhang, Dong Yu, InterSpeech 2020

Citations

@inproceedings{Weng2020,
  author={Chao Weng and Chengzhu Yu and Jia Cui and Chunlei Zhang and Dong Yu},
  title={{Minimum Bayes Risk Training of RNN-Transducer for End-to-End Speech Recognition}},
  year=2020,
  booktitle={Proc. Interspeech 2020},
  pages={966--970},
  doi={10.21437/Interspeech.2020-1221},
  url={http://dx.doi.org/10.21437/Interspeech.2020-1221}
}

@inproceedings{Weng2018,
  author={Chao Weng and Jia Cui and Guangsen Wang and Jun Wang and Chengzhu Yu and Dan Su and Dong Yu},
  title={Improving Attention Based Sequence-to-Sequence Models for End-to-End English Conversational Speech Recognition},
  year=2018,
  booktitle={Proc. Interspeech 2018},
  pages={761--765},
  doi={10.21437/Interspeech.2018-1030},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1030}
}

Disclaimer

This is not an officially supported Tencent product