WenetSpeech Dataset¶
A Mandarin speech dataset with over 10,000 hours of audio, usage link: PPASR
The WenetSpeech Dataset contains over 10,000 hours of Mandarin speech data, all sourced from YouTube and Podcasts. OCR (Optical Character Recognition) and ASR (Automatic Speech Recognition) technologies are used to label each YouTube and Podcast recording. To improve corpus quality, WenetSpeech employs an innovative end-to-end label error detection method to further validate and filter the data.
- All data is classified into 3 categories as shown in the table below:
| Data Category | Duration (hours) | Confidence | Applicable Systems |
|---|---|---|---|
| Strongly Labeled | 10005 | >=0.95 | Supervised Training |
| Weakly Labeled | 2478 | [0.6, 0.95] | Semi-supervised or Noisy Training |
| Unlabeled | 9952 | / | Unsupervised Training or Pretraining |
| Total | 22435 | / | / |
- Domains, speaking styles, and scenarios of the strongly labeled data are divided into 10 groups as shown below:
| Domain | YouTube (hours) | Podcast (hours) | Total (hours) |
|---|---|---|---|
| Audiobooks | 0 | 250.9 | 250.9 |
| Live Commentary | 112.6 | 135.7 | 248.3 |
| Documentary | 386.7 | 90.5 | 477.2 |
| Drama | 4338.2 | 0 | 4338.2 |
| Interviews | 324.2 | 614 | 938.2 |
| News | 0 | 868 | 868 |
| Reading | 0 | 1110.2 | 1110.2 |
| Discussion | 204 | 90.7 | 294.7 |
| Variety Shows | 603.3 | 224.5 | 827.8 |
| Others | 144 | 507.5 | 651.5 |
| Total | 6113 | 3892 | 10005 |
- Three subsets (S, M, L) for building ASR systems with different data scales:
| Training Data | Confidence | Duration (hours) |
|---|---|---|
| L | [0.95, 1.0] | 10005 |
| M | 1.0 | 1000 |
| S | 1.0 | 100 |
- Evaluation and test data:
| Evaluation Data | Duration (hours) | Source | Description |
|---|---|---|---|
| DEV | 20 | Internet | Designed for cross-validation in speech tools during training |
| TEST_NET | 23 | Internet | Competition test set |
| TEST_MEETING | 15 | Meetings | Far-field, dialogue, spontaneous, and meeting datasets |
1. Tutorial: Training an ASR Model with WenetSpeech¶
This tutorial introduces how to train a speech recognition model using only the strongly labeled data, mainly in three steps:
-
Download and Extract the Dataset
Download the dataset from the official website after filling out the form. You will receive an email with download instructions. Execute the three commands in the email to download and extract the dataset. Note: This requires 500GB of disk space. -
Prepare the Dataset
The downloaded data is unprocessed. Use thecreate_wenetspeech_data.pyscript in thetoolsdirectory to segment and label the audio files. This step requires 3TB of disk space.
cd tools/
python create_wenetspeech_data.py --wenetspeech_json=/media/wenetspeech/WenetSpeech.json
- Create Training Data
Similar to standard data preparation, runcreate_data.pyin the project root directory to generate data lists, vocabulary files, and mean/std files for training. After this step, you can start training the model (see Training Model for details).
python create_data.py
Project address: https://github.com/yeyupiaoling/PPASR