|
File : README
|
|
|
|
|
|
|
|
AISHELL-2 Speech Data
|
|
|
|
北京希尔贝壳科技有限公司
|
|
Beijing Shell Shell Technology Co.,Ltd
|
|
03/10/2018
|
|
|
|
1. AISHELL-2 Speech Data
|
|
|
|
- Sampling Rate : 16kHz
|
|
- Sample Format : 16bit
|
|
- Environment : Indoor
|
|
- Speech Data Type : PCM
|
|
- Channel Number : 1
|
|
- Recording Equipment : iOS
|
|
- Time : 1000 hours of speech data (around 1 million utterances)
|
|
- Speaker : 1991 speakers (845 male and 1146 female)
|
|
|
|
|
|
2. 数据目录结构
|
|
│ AISHELL2-Data-Specification .docx (Corpus Information)
|
|
│ README.txt (readme)
|
|
│ ChangeLog (Change Information)
|
|
└─ data (Data File)
|
|
│├─wav.list (Audio list)
|
|
│├─trans.txt (Transcript Content)
|
|
│├─spk_info.txt (Speaker Information)
|
|
│├─wav (Audio Data File)
|
|
│├─D2166 (Speaker ID File)
|
|
││ ││ ID2166W0001.wav (Audio)
|
|
|
|
3. Evaluation Data:
|
|
Currently we open-sourced AISHELL2-2018A-EVAL
|
|
- dev: 2500 utterances from 5 speaker
|
|
- test: 5000 utterances from 10 speakers
|
|
You can download above evaluation set from: https://www.aishelltech.com/aishell_2018_eval
|
|
|
|
4. RECIPE
|
|
Based on Kaldi standard system, AISHELL-2 provides a self-contained Mandarin ASR recipe
|
|
- a word segmentation module, which is a must-have component for Chinese ASR systems
|
|
- an open-sourced Mandarin lexicon(DaCiDian)
|
|
- a simplified GMM training recipe
|
|
- 80-dim FBank without pitch as NN input feature
|
|
- slightly different data augmentation setup(tempo perturbation, max-volume perturbation)
|
|
- acoustic channel adaptation recipe(AM fine-tuning)
|
|
- a real-time streaming ASR demo via laptop microphone (tested on MacOS)
|
|
You can download recipe set from: http://www.aishelltech.com/aishell_2
|
|
|