mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-26 10:16:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

added results on zh-HK

Browse Source
This commit is contained in:
jinzr 2024-03-20 12:39:21 +08:00
parent 6993183dd7
commit 4237127be2
1 changed files with 74 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

83
egs/commonvoice/ASR/RESULTS.md
View File

@ -1,4 +1,73 @@
## Results
### Commonvoice Cantonese (zh-HK) Char training results (Zipformer)
See #1542 for more details.
Number of model parameters: 72526519, i.e., 72.53 M
The best CER, for CommonVoice 16.1 (cv-corpus-16.1-2023-12-06/zh-HK) is below:
| | Dev | Test | Note |
|----------------------|-------|------|--------------------|
| greedy_search | 1.17 | 1.22 | --epoch 24 --avg 5 |
| modified_beam_search | 0.98 | 1.11 | --epoch 24 --avg 5 |
| fast_beam_search | 1.08 | 1.27 | --epoch 24 --avg 5 |
When doing the cross-corpus validation on MDCC (w/o blank penalty),
the best CER is below:
| | Dev | Test | Note |
|----------------------|-------|------|--------------------|
| greedy_search | 42.40 | 42.03| --epoch 24 --avg 5 |
| modified_beam_search | 39.73 | 39.19| --epoch 24 --avg 5 |
| fast_beam_search | 42.14 | 41.98| --epoch 24 --avg 5 |
When doing the cross-corpus validation on MDCC (with blank penalty set to 2.2),
the best CER is below:
| | Dev | Test | Note |
|----------------------|-------|------|----------------------------------------|
| greedy_search | 39.19 | 39.09| --epoch 24 --avg 5 --blank-penalty 2.2 |
| modified_beam_search | 37.73 | 37.65| --epoch 24 --avg 5 --blank-penalty 2.2 |
| fast_beam_search | 37.73 | 37.74| --epoch 24 --avg 5 --blank-penalty 2.2 |
To reproduce the above result, use the following commands for training:
```bash
export CUDA_VISIBLE_DEVICES="0,1"
./zipformer/train_char.py \
--world-size 2 \
--num-epochs 30 \
--start-epoch 1 \
--use-fp16 1 \
--exp-dir zipformer/exp \
--cv-manifest-dir data/zh-HK/fbank \
--language zh-HK \
--use-validated-set 1 \
--context-size 1 \
--max-duration 1000
```
and the following commands for decoding:
```bash
for method in greedy_search modified_beam_search fast_beam_search; do
./zipformer/decode_char.py \
--epoch 24 \
--avg 5 \
--decoding-method $method \
--exp-dir zipformer/exp \
--cv-manifest-dir data/zh-HK/fbank \
--context-size 1 \
--language zh-HK
done
```
Detailed experimental results and pre-trained model are available at:
<https://huggingface.co/zrjin/icefall-asr-commonvoice-zh-HK-zipformer-2024-03-20>
### GigaSpeech BPE training results (Pruned Stateless Transducer 7)
#### [pruned_transducer_stateless7](./pruned_transducer_stateless7)
@ -13,8 +82,8 @@ Results are:
| | Dev | Test |
|----------------------|-------|-------|
| greedy search | 9.96 | 12.54 |
| modified beam search | 9.86 | 12.48 |
| greedy_search | 9.96 | 12.54 |
| modified_beam_search | 9.86 | 12.48 |
To reproduce the above result, use the following commands for training:
@ -55,10 +124,6 @@ and the following commands for decoding:
Pretrained model is available at
<https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17>
The tensorboard log for training is available at
<https://tensorboard.dev/experiment/j4pJQty6RMOkMJtRySREKw/>
### Commonvoice (fr) BPE training results (Pruned Stateless Transducer 7_streaming)
#### [pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
@ -73,9 +138,9 @@ Results are:
| decoding method | Test |
|----------------------|-------|
| greedy search | 9.95 |
| modified beam search | 9.57 |
| fast beam search | 9.67 |
| greedy_search | 9.95 |
| modified_beam_search | 9.57 |
| fast_beam_search | 9.67 |
Note: This best result is trained on the full librispeech and gigaspeech, and then fine-tuned on the full commonvoice.