k2_zipformer2_english_v1
Zipformer2 recipe derived from : https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/zipformer
Training data : CommonVoice, VoxPopuli (speed-perturb 3x):
- cca 1500 + 3x500 hours of training data
Output text/symbols include:
- TrueCase capitalization
- punctuation
[,.?!]as standalone tokens
Config:
--num-epochs 20 \
--base-lr 0.04 \
\
--causal 1 \
--use-transducer 1 \
--use-ctc 0 \
\
--num-encoder-layers 2,2,2,2,2,2 \
--feedforward-dim 512,768,768,768,768,768 \
--encoder-dim 192,256,256,256,256,256 \
--encoder-unmasked-dim 192,192,192,192,192,192 \
Results
| ID | System | cv-dev | cv-test | vp-dev | vp-test | Comment |
|---|---|---|---|---|---|---|
| A | small (24M) | 18.57 | 21.98 | 13.66 | 13.26 | ep20,avg4 |
- non-streaming results from decode.py
- cv = CommonVoice, vp = VoxPopuli
- exported the model averaging : ep20,avg4
Note
- Not the best possible results, this model is for integration tests