Meta-Transfer Learning for Code-Switched Speech Recognition

Genta Indra Winata, Samuel Cahyawijaya, Zhaojiang Lin, Zihan Liu, Peng Xu, Pascale Fung

This is the implementation of our paper accepted in ACL 2020.

This code has been written using PyTorch. If you use any source codes or datasets included in this toolkit in your work, please cite the following paper.

@inproceedings{winata-etal-2020-meta,
    title = "Meta-Transfer Learning for Code-Switched Speech Recognition",
    author = "Winata, Genta Indra  and
      Cahyawijaya, Samuel  and
      Lin, Zhaojiang  and
      Liu, Zihan  and
      Xu, Peng  and
      Fung, Pascale",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.348",
    pages = "3770--3776",
}

Abstract

An increasing number of people in the world today speak a mixed-language as a result of being multilingual. However, building a speech recognition system for code-switching remains difficult due to the availability of limited resources and the expense and significant effort required to collect mixed-language data. We therefore propose a new learning method, meta-transfer learning, to transfer learn on a code-switched speech recognition system in a low-resource setting by judiciously extracting information from high-resource monolingual datasets. Our model learns to recognize individual languages, and transfer them so as to better recognize mixed-language speech by conditioning the optimization on the code-switching data. Based on experimental results, our model outperforms existing baselines on speech recognition and language modeling tasks, and is faster to converge.

Data

• SEAME Phase II datasets (https://catalog.ldc.upenn.edu/LDC2015S04)
• HKUST (https://catalog.ldc.upenn.edu/LDC2005S15)
• CommonVoice v2 (Jun 2019) Kindly check the /data/ directory to check the data split and labels. This is the direct link (https://voice-prod-bundler-ee1969a6ce8178826482b88e843c335139bd3fb4.s3.amazonaws.com/cv-corpus-2/en.tar.gz)

Model Architecture

Setup

• Install PyTorch (Tested in PyTorch 1.0 and Python 3.6)
• Install library dependencies (requirement.txt)

Run the code

• Meta-Transfer Learning

python meta_transfer_train.py --train-manifest-list data/manifests/cv-valid-train_manifest.csv data/manifests/hkust_16khz_train_manifest.csv data/manifests/seame_phaseII_train_manifest.csv \
--train-partition-list 1 1 1 --valid-manifest-list data/manifests/cv-valid-dev_manifest.csv data/manifests/hkust_16khz_dev_manifest.csv data/manifests/seame_phaseII_val_manifest.csv \ 
--name mtl_enc2_dec4_512_b8_16khz_copy_grad --cuda --k-train 8 --k-valid 8 --labels-path data/labels/hkust_seame_labels.json --lr 1e-4 --save-folder save/ --save-every 10000 \ 
--feat_extractor vgg_cnn --dropout 0.1 --num-enc-layers 2 --num-dec-layers 4 --num-heads 8 --dim-model 512 --dim-key 64 --dim-value 64 --dim-input 5120 --dim-inner 512 --dim-emb 512 --early-stop cer,200 \
--src-max-len 5000 --tgt-max-len 2500 --evaluate-every 10000 --epochs 1000000 --sample-rate 16000 --copy-grad

• Joint training

python joint_train.py --train-manifest-list data/manifests/cv-valid-train_manifest.csv data/manifests/hkust_16khz_train_manifest.csv data/manifests/seame_phaseII_train_manifest.csv \
--valid-manifest-list data/manifests/cv-valid-dev_manifest.csv data/manifests/hkust_16khz_dev_manifest.csv data/manifests/seame_phaseII_val_manifest.csv --cuda --k-train 8 \
--labels-path data/labels/hkust_seame_labels.json --lr 1e-4 --name joint_enc2_dec4_512_b8_16khz --save-folder save/ --save-every 10000 --feat_extractor vgg_cnn --dropout 0.1 --num-enc-layers 2 \
--num-dec-layers 4 --num-heads 8 --dim-model 512 --dim-key 64 --dim-value 64 --dim-input 5120 --dim-inner 512 --dim-emb 512 --early-stop cer,200 --src-max-len 5000 --tgt-max-len 2500 --evaluate-every 10000 \
--epochs 10000000 --sample-rate 16000 --train-partition-list 1 1 1

Bug Report

Feel free to create an issue or send email to [email protected]