WenetSpeech
A 10000+ Hours Multi-domain Chinese Corpus for Speech Recognition
Download
Please visit the official website, read the license, and follow the instruction to download the data.
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
- WenetSpeech refers a lot of work of GigaSpeech, and we thank Jiayu Du and Guoguo Chen for their suggestions on this work.
- We thank Xi'an Future AI Innovation Center for providing hosting service for WenetSpeech. We also thank MindSpore for the support of this work, which is a new deep learning computing framework.
- Our gratitude goes to Lianhui Zhang and Yu Mao for collecting some of the YouTube data.
6 Jan 03, 2023
99 Dec 09, 2022
17 Sep 25, 2022
958 Jan 03, 2023
11 Sep 05, 2022
1 Jan 14, 2022
75 Dec 28, 2022
125 Jan 03, 2023
67 Dec 21, 2022
93 Sep 30, 2022
3 Jul 10, 2022
36 Dec 26, 2022
23 Oct 19, 2022
121 Dec 30, 2022
17 Aug 16, 2022
3 May 09, 2022
85 Nov 27, 2022
3 Jan 04, 2023
11 Oct 17, 2022
1.1k Jan 06, 2023