Add small streaming Zipformer transducer model (#903)

This commit is contained in:
Desh Raj 2023-02-13 02:53:28 -05:00 committed by GitHub
parent c34ee67691
commit 6a8b649e56
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -77,6 +77,18 @@ for m in greedy_search modified_beam_search fast_beam_search; do
done done
``` ```
#### Smaller model
A smaller model (~20M params) is also available with configuration based on [this comment](https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740). The WERs are:
| decoding method | chunk size | test-clean | test-other | comment | decoding mode |
|----------------------|------------|------------|------------|---------------------|----------------------|
| greedy search | 320ms | 3.94 | 9.79 | --epoch 30 --avg 9 | simulated streaming |
| modified beam search | 320ms | 3.88 | 9.53 | --epoch 30 --avg 9 | simulated streaming |
You can find a pretrained model, training logs, decoding logs, and decoding
results at: <https://huggingface.co/desh2608/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-small>
### zipformer_mmi (zipformer with mmi loss) ### zipformer_mmi (zipformer with mmi loss)