The WER comparison between current benchmark and DS2 paper
Created by: kuke
With the latest update, the BaiduEN8K model has caught up with the original Deep Speech 2 work in WER performance on some public test datasets.
| On pubic LM (8.3G) | On internal LM (260G) | DS2 paper | |
|---|---|---|---|
| LibriSpeech Test-Clean | 5.41 | 5.28 | 5.33 |
| LibriSpeech Test-Other | 13.85 | 13.49 | 13.26 |
| VoxForge American-Canadian | 7.13 | 6.96 | 7.55 |
| VoxForge Commonwealth | 14.93 | 14.62 | 13.56 |
| VoxForge European | 18.64 | 18.34 | 17.55 |
| VoxForge Indian | 25.51 | 25.27 | 22.44 |
- Training data set: 8628h vs. 11940h