wenet-e2e / WenetSpeech

A 10000+ hours dataset for Chinese speech recognition
Apache License 2.0
506 stars 49 forks source link

WenetSpeech

Official website | Paper

A 10000+ Hours Multi-domain Chinese Corpus for Speech Recognition

WenetSpeech

Download

Please visit the official website, read the license, and follow the instruction to apply for the PASSWORD to download the data.

echo 'PASSWORD' > SAFEBOX/password

From Tecent Meeting (default)

Download WenetSpeech:

bash utils/download_wenetspeech.sh DOWNLOAD_DIR UNTAR_DIR

From ModelScope

Install modelscope (depends on torch) before downloading:

conda create -n modelscope python=3.7
conda activate modelscope
pip install torch
pip install modelscope -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html

Download WenetSpeech from modelscope:

sed -i 's/modelscope=false/modelscope=true/g' utils/download_wenetspeech.sh
bash utils/download_wenetspeech.sh DOWNLOAD_DIR UNTAR_DIR

Discussion & Communication

Please scan the QR code on the left to follow our offical account of WeNet. We created a WeChat group for better discussion and quicker response. Please scan the personal QR code on the right, and the guy is responsible for inviting you to the chat group.

Benchmark

Toolkit Dev Test_Net Test_Meeting AIShell-1
Kaldi 9.07 12.83 24.72 5.41
ESPNet 9.70 8.90 15.90 3.90
WeNet 8.88 9.70 15.59 4.61

Description

Creation

All the data are collected from YouTube and Podcast. Optical character recognition (OCR) and automatic speech recognition (ASR) techniques are adopted to label each YouTube and Podcast recording, respectively. To improve the quality of the corpus, we use a novel end-to-end label error detection method to further validate and filter the data.

Categories

In summary, WenetSpeech groups all data into 3 categories, as the following table shows:

Set Hours Confidence Usage
High Label 10005 >=0.95 Supervised Training
Weak Label 2478 [0.6, 0.95] Semi-supervised or noise training
Unlabel 9952 / Unsupervised training or Pre-training
In Total 22435 / All above

High Label Data

We classify the high label into 10 groups according to its domain, speaking style, and scenarios.

Domain Youtube Podcast Total
audiobook 0 250.9 250.9
commentary 112.6 135.7 248.3
documentary 386.7 90.5 477.2
drama 4338.2 0 4338.2
interview 324.2 614 938.2
news 0 868 868
reading 0 1110.2 1110.2
talk 204 90.7 294.7
variety 603.3 224.5 827.8
others 144 507.5 651.5
Total 6113 3892 10005

As shown in the following table, we provide 3 training subsets, namely S, M and L for building ASR systems on different data scales.

Training Subsets Confidence Hours
L [0.95, 1.0] 10005
M 1.0 1000
S 1.0 100

Evaluation Sets

Evaluation Sets Hours Source Description
DEV 20 Internet Specially designed for some speech tools which require cross-validation set in training
TEST_NET 23 Internet Match test
TEST_MEETING 15 Real meeting Mismatch test which is a far-field, conversational, spontaneous, and meeting dataset

Contributors

ACKNOWLEDGEMENTS