mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-09 05:55:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update training commands and decode.py accuracy values, add streaming model section

Browse Source
This commit is contained in:
Kinan Martin 2025-09-03 17:54:34 +09:00
parent ef7664e7cf
commit bc2560cb7a
1 changed files with 82 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

99
egs/multi_ja_en/ASR/RESULTS.md
View File

@ -2,29 +2,32 @@
### Zipformer
#### Non-streaming
#### Non-streaming (Byte-Level BPE vocab_size=2000)
Trained on 15k hours of ReazonSpeech (filtered to only audio segments between 8s and 22s) and 15k hours of MLS English.
The training command is:
```shell
./zipformer/train.py \
--world-size 4 \
--num-epochs 21 \
--world-size 8 \
--num-epochs 10 \
--start-epoch 1 \
--use-fp16 1 \
--exp-dir zipformer/exp \
--manifest-dir data/manifests
--manifest-dir data/manifests \
--enable-musan True
```
The decoding command is:
```shell
./zipformer/decode.py \
--epoch 21 \
--avg 15 \
--epoch 10 \
--avg 1 \
--exp-dir ./zipformer/exp \
--max-duration 600 \
--decoding-method greedy_search
--decoding-method modified_beam_search \
--manifest-dir data/manifests
```
To export the model with onnx:
@ -33,28 +36,28 @@ To export the model with onnx:
./zipformer/export-onnx.py \
--tokens ./data/lang/bbpe_2000/tokens.txt \
--use-averaged-model 0 \
--epoch 21 \
--epoch 10 \
--avg 1 \
--exp-dir ./zipformer/exp
```
Word Error Rates (WERs) listed below:
WER and CER on test set listed below (calculated with `./zipformer/decode.py`):
| Datasets | ReazonSpeech | ReazonSpeech | LibriSpeech | LibriSpeech |
|----------------------|--------------|---------------|--------------------|-------------------|
| Zipformer WER (%) | dev | test | test-clean | test-other |
| greedy_search | 5.9 | 4.07 | 3.46 | 8.35 |
| modified_beam_search | 4.87 | 3.61 | 3.28 | 8.07 |
| Datasets | ReazonSpeech + MLS English (combined test set) |
|----------------------|------------------------------------------------|
| Zipformer WER (%) | test |
| greedy_search | 6.33 |
| modified_beam_search | 6.32 |
We also include WER% for common English ASR datasets:
| Corpus | WER (%) |
| Corpus | WER (%) |
|-----------------------------|---------|
| CommonVoice | 29.03 |
| TED | 16.78 |
| MLS English (test-clean) | 8.64 |
| MLS English (test set) | 8.64 |
And CER% for common Japanese datasets:
@ -68,3 +71,65 @@ And CER% for common Japanese datasets:
Pre-trained model can be found here: [https://huggingface.co/reazon-research/reazonspeech-k2-v2-ja-en/tree/multi_ja_en_15k15k](https://huggingface.co/reazon-research/reazonspeech-k2-v2-ja-en/tree/multi_ja_en_15k15k)
(Not yet publicly released)
#### Streaming (Byte-Level BPE vocab_size=2000)
Trained on 15k hours of ReazonSpeech (filtered to only audio segments between 8s and 22s) and 15k hours of MLS English.
The training command is:
```shell
./zipformer/train.py \
--world-size 8 \
--num-epochs 10 \
--start-epoch 1 \
--use-fp16 1 \
--exp-dir zipformer/exp \
--manifest-dir data/manifests \
--enable-musan True
```
The decoding command is:
```shell
./zipformer/decode.py \
--epoch 10 \
--avg 1 \
--exp-dir ./zipformer/exp \
--decoding-method modified_beam_search \
--manifest-dir data/manifests
```
To export the model with onnx:
```shell
./zipformer/export-onnx.py \
--tokens ./data/lang/bbpe_2000/tokens.txt \
--use-averaged-model 0 \
--epoch 10 \
--avg 1 \
--decode-chunk-len 32 \
--exp-dir ./zipformer/exp
```
You may also use decode chunk sizes `16`, `32`, `64`, `128`.
Word Error Rates (WERs) listed below:
*Please let us know which script to use to evaluate the streaming model!*
We also include WER% for common English ASR datasets:
*Please let us know which script to use to evaluate the streaming model!*
And CER% for common Japanese datasets:
*Please let us know which script to use to evaluate the streaming model!*
Pre-trained model can be found here: [https://huggingface.co/reazon-research/reazonspeech-k2-v2-ja-en/tree/multi_ja_en_15k15k](https://huggingface.co/reazon-research/reazonspeech-k2-v2-ja-en/tree/multi_ja_en_15k15k)
(Not yet publicly released)