mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-09 18:12:19 +00:00
0855b0338a
* add streaming support to reazonresearch * update README for streaming * Update RESULTS.md * add onnx decode --------- Co-authored-by: root <root@KDA03.cm.cluster> Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com> Co-authored-by: root <root@KDA01.cm.cluster> Co-authored-by: zr_jin <peter.jin.cn@gmail.com>
88 lines
2.3 KiB
Markdown
88 lines
2.3 KiB
Markdown
## Results
|
|
|
|
### Zipformer
|
|
|
|
#### Non-streaming
|
|
|
|
##### large-scaled model, number of model parameters: 159337842, i.e., 159.34 M
|
|
|
|
| decoding method | In-Distribution CER | JSUT | CommonVoice | TEDx | comment |
|
|
| :------------------: | :-----------------: | :--: | :---------: | :---: | :----------------: |
|
|
| greedy search | 4.2 | 6.7 | 7.84 | 17.9 | --epoch 39 --avg 7 |
|
|
| modified beam search | 4.13 | 6.77 | 7.69 | 17.82 | --epoch 39 --avg 7 |
|
|
|
|
The training command is:
|
|
|
|
```shell
|
|
./zipformer/train.py \
|
|
--world-size 8 \
|
|
--num-epochs 40 \
|
|
--start-epoch 1 \
|
|
--use-fp16 1 \
|
|
--exp-dir zipformer/exp-large \
|
|
--causal 0 \
|
|
--num-encoder-layers 2,2,4,5,4,2 \
|
|
--feedforward-dim 512,768,1536,2048,1536,768 \
|
|
--encoder-dim 192,256,512,768,512,256 \
|
|
--encoder-unmasked-dim 192,192,256,320,256,192 \
|
|
--lang data/lang_char \
|
|
--max-duration 1600
|
|
```
|
|
|
|
The decoding command is:
|
|
|
|
```shell
|
|
./zipformer/decode.py \
|
|
--epoch 40 \
|
|
--avg 16 \
|
|
--exp-dir zipformer/exp-large \
|
|
--max-duration 600 \
|
|
--causal 0 \
|
|
--decoding-method greedy_search \
|
|
--num-encoder-layers 2,2,4,5,4,2 \
|
|
--feedforward-dim 512,768,1536,2048,1536,768 \
|
|
--encoder-dim 192,256,512,768,512,256 \
|
|
--encoder-unmasked-dim 192,192,256,320,256,192 \
|
|
--lang data/lang_char \
|
|
--blank-penalty 0
|
|
```
|
|
|
|
#### Streaming
|
|
|
|
We have not completed evaluation of our models yet and will add evaluation results here once it's completed.
|
|
|
|
The training command is:
|
|
```shell
|
|
./zipformer/train.py \
|
|
--world-size 8 \
|
|
--num-epochs 40 \
|
|
--start-epoch 1 \
|
|
--use-fp16 1 \
|
|
--exp-dir zipformer/exp-large \
|
|
--causal 1 \
|
|
--num-encoder-layers 2,2,4,5,4,2 \
|
|
--feedforward-dim 512,768,1536,2048,1536,768 \
|
|
--encoder-dim 192,256,512,768,512,256 \
|
|
--encoder-unmasked-dim 192,192,256,320,256,192 \
|
|
--lang data/lang_char \
|
|
--max-duration 1600
|
|
```
|
|
|
|
The decoding command is:
|
|
|
|
```shell
|
|
./zipformer/streaming_decode.py \
|
|
--epoch 28 \
|
|
--avg 15 \
|
|
--causal 1 \
|
|
--chunk-size 32 \
|
|
--left-context-frames 256 \
|
|
--exp-dir ./zipformer/exp-large \
|
|
--lang data/lang_char \
|
|
--num-encoder-layers 2,2,4,5,4,2 \
|
|
--feedforward-dim 512,768,1536,2048,1536,768 \
|
|
--encoder-dim 192,256,512,768,512,256 \
|
|
--encoder-unmasked-dim 192,192,256,320,256,192
|
|
```
|
|
|