Results on Switchboard datasets.
- Reported in "Advancing CTC-CRF Based End-to-End Speech Recognition with Wordpieces and Conformers"
- AM: Conformer with 52M parameters. SpecAug and 3-way perturbation is applied.
- "Trans." in the table denotes the interpolation between 4-gram and Transformer LM.
- Data for phone-based system and wp-based system rescoring respectively is publicly available on Google Drive, including
data/lang_{phn,bpe},Nbest list.
| Unit | LM | SW | CH | Eval2000 | Notes |
|---|---|---|---|---|---|
| phone | 4-gram | 7.9 | 16.1 | 12.1 | --- |
| phone | Trans. | 6.9 | 14.5 | 10.7 | N-best rescoring, N=40, weight=0.8 |
| wp | 4-gram | 8.7 | 16.5 | 12.7 | --- |
| wp | Trans. | 7.2 | 14.8 | 11.1 | N-best rescoring, N=60, weight=0.8 |
Experiment
-
Phone-based system
bash run.sh
-
Char-based or wordpiece-based
bash run_wp.sh
For rescoring with "Trans.", please refer to local/pytorchnn/readme.
- AM: VGG-BLSTM with 16M parameters. 3-way perturbation is applied.
- Hyper-parameters of AM training:
lamb=0.1, hdim=320, lr=0.001
| Unit | SW | CH |
|---|---|---|
| phone | 9.9 | 19.4 |
- AM: BLSTM with 13M parameters.
- Hyper-parameters of AM training:
lamb=0.1, hdim=320, lr=0.001
| Unit | SW | CH | SP |
|---|---|---|---|
| phone | 11.0 | 21.0 | N |
| phone | 10.3 | 19.7 | Y |
| char | 12.7 | 24.0 | N |
| char | 11.4 | 21.7 | Y |