mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 10:02:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update RESULTS.md, adding results and model links of zipformer-small/medium CTC/AED models (#1683)

Browse Source
This commit is contained in:
Zengwei Yao 2024-07-09 20:04:47 +08:00 committed by GitHub
parent 1c3d992a39
commit 785f3f0bcf
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 112 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
egs/librispeech/ASR/README.md
View File

@ -50,7 +50,7 @@ We place an additional Conv1d layer right after the input embedding layer.
| `conformer-ctc2` | Reworked Conformer | Use auxiliary attention head |
| `conformer-ctc3` | Reworked Conformer | Streaming version + delay penalty |
| `zipformer-ctc` | Zipformer | Use auxiliary attention head |
| `zipformer` | Upgraded Zipformer | Use auxiliary transducer head | The latest recipe |
| `zipformer` | Upgraded Zipformer | Use auxiliary transducer head / attention-decoder head | The latest recipe |
# MMI

113
egs/librispeech/ASR/RESULTS.md
View File

@ -8,6 +8,117 @@ See <https://github.com/k2-fsa/icefall/pull/1389> for more details.
#### Non-streaming
##### small-scale model, number of model parameters: 46282107, i.e., 46.3 M
You can find a pretrained model, training logs, decoding logs, and decoding results at:
<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-small-ctc-attention-decoder-2024-07-09>
You can use <https://github.com/k2-fsa/sherpa> to deploy it.
| decoding method | test-clean | test-other | comment |
|--------------------------------------|------------|------------|---------------------|
| ctc-decoding | 3.04 | 7.04 | --epoch 50 --avg 30 |
| attention-decoder-rescoring-no-ngram | 2.45 | 6.08 | --epoch 50 --avg 30 |
The training command is:
```bash
export CUDA_VISIBLE_DEVICES="0,1"
# For non-streaming model training:
./zipformer/train.py \
--world-size 2 \
--num-epochs 50 \
--start-epoch 1 \
--use-fp16 1 \
--exp-dir zipformer/exp-small \
--full-libri 1 \
--use-ctc 1 \
--use-transducer 0 \
--use-attention-decoder 1 \
--ctc-loss-scale 0.1 \
--attention-decoder-loss-scale 0.9 \
--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 \
--base-lr 0.04 \
--max-duration 1700 \
--master-port 12345
```
The decoding command is:
```bash
export CUDA_VISIBLE_DEVICES="0"
for m in ctc-decoding attention-decoder-rescoring-no-ngram; do
./zipformer/ctc_decode.py \
--epoch 50 \
--avg 30 \
--exp-dir zipformer/exp-small \
--use-ctc 1 \
--use-transducer 0 \
--use-attention-decoder 1 \
--attention-decoder-loss-scale 0.9 \
--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 \
--max-duration 100 \
--causal 0 \
--num-paths 100 \
--decoding-method $m
done
```
##### medium-scale model, number of model parameters: 89987295, i.e., 90.0 M
You can find a pretrained model, training logs, decoding logs, and decoding results at:
<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-ctc-attention-decoder-2024-07-08>
You can use <https://github.com/k2-fsa/sherpa> to deploy it.
| decoding method | test-clean | test-other | comment |
|--------------------------------------|------------|------------|---------------------|
| ctc-decoding | 2.46 | 5.57 | --epoch 50 --avg 22 |
| attention-decoder-rescoring-no-ngram | 2.23 | 4.98 | --epoch 50 --avg 22 |
The training command is:
```bash
export CUDA_VISIBLE_DEVICES="0,1,2,3"
# For non-streaming model training:
./zipformer/train.py \
--world-size 4 \
--num-epochs 50 \
--start-epoch 1 \
--use-fp16 1 \
--exp-dir zipformer/exp \
--full-libri 1 \
--use-ctc 1 \
--use-transducer 0 \
--use-attention-decoder 1 \
--ctc-loss-scale 0.1 \
--attention-decoder-loss-scale 0.9 \
--max-duration 1200 \
--master-port 12345
```
The decoding command is:
```bash
export CUDA_VISIBLE_DEVICES="0"
for m in ctc-decoding attention-decoder-rescoring-no-ngram; do
./zipformer/ctc_decode.py \
--epoch 50 \
--avg 22 \
--exp-dir zipformer/exp \
--use-ctc 1 \
--use-transducer 0 \
--use-attention-decoder 1 \
--attention-decoder-loss-scale 0.9 \
--max-duration 100 \
--causal 0 \
--num-paths 100 \
--decoding-method $m
done
```
##### large-scale model, number of model parameters: 174319650, i.e., 174.3 M
You can find a pretrained model, training logs, decoding logs, and decoding results at:
@ -15,8 +126,6 @@ You can find a pretrained model, training logs, decoding logs, and decoding resu
You can use <https://github.com/k2-fsa/sherpa> to deploy it.
Results of the CTC head:
| decoding method | test-clean | test-other | comment |
|--------------------------------------|------------|------------|---------------------|
| ctc-decoding | 2.29 | 5.14 | --epoch 50 --avg 29 |