Implementation of Baidu Warp-CTC using torch7. Creates a network based on the DeepSpeech2 architecture using the Torch7 library, trained with the CTC activation function.
- Train large models with large datasets via online loading using LMDB and multi-GPU support.
- Supports variable length batches via masking.
- Implements the AN4 Audio database (50 mins of data). Has also been extended to train using the LibriSpeech dataset (1000 hours of data). Custom dataset preparation is explained in documentation.
There are currently two branches, Master and Phoneme:
- Master: This branch trains DeepSpeech2. Also included is an evaluation script which calculates the WER/CER, as well as a prediction script. This branch is useful for understanding how the DeepSpeech and CTC works and is easy to run after installation. Highly recommended to checkout this branch.
- Phonemes: This branch is experimental and uses phonemes rather than character based predictions. This is fully credited and extended by CCorfield and his awesome work in porting to use phonemes. In addition to this I'd like to also thank Shane Walker for his awesome recent conversion to use phonemes as well.
These results are based on training on the AN4 training set, and testing on the AN4 test set. Will be updated as architecture/datasets changes.
| WER | CER |
|---|---|
| 14 | 4.22 |
Follow Instructions/Data Preparation/Documentation found in the wiki here to set up and run the code.
Technical documentation can be found here.
Lots of people helped/contributed to this project that deserve recognition:
- Soumith Chintala for his support on Torch7 and the vast open source projects he has contributed that made this project possible!
- Charles Corfield for his work on the Phoneme Dataset and his overall contribution and aid throughout.
- Will Frey for his thorough communication and aid in the development process.
- Ding Ling, Yuan Yang and Yan Xia for their significant contribution to online training, multi-gpu support and many other important features.
- Erich Elsen and the team from Baidu for their contribution of Warp-CTC that made this possible, and the encouraging words and support given throughout the project.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent