mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-07 16:14:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update results

Browse Source
This commit is contained in:
Guanbo Wang 2022-04-11 21:46:40 +00:00
parent f485b66d54
commit 22f011e5ab
2 changed files with 35 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -15,6 +15,6 @@ ln -sfv /path/to/GigaSpeech download/GigaSpeech
## Performance Record ## Performance Record
| | Dev | Test | | | Dev | Test |
|-----|-------|-------| |-----|-------|-------|
| WER | 11.93 | 11.86 | | WER | 10.47 | 10.58 |
See [RESULTS](/egs/gigaspeech/ASR/RESULTS.md) for details. See [RESULTS](/egs/gigaspeech/ASR/RESULTS.md) for details.

41
egs/gigaspeech/ASR/RESULTS.md
View File

@ -5,22 +5,23 @@
#### 2022-04-06 #### 2022-04-06
The best WER, as of 2022-04-06, for the gigaspeech is below The best WER, as of 2022-04-06, for the gigaspeech is below
(using HLG decoding + n-gram LM rescoring + attention decoder rescoring):
Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
| | Dev | Test | | | Dev | Test |
|-----|-------|-------| |-----|-------|-------|
| WER | 11.93 | 11.86 | | WER | 10.47 | 10.58 |
Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are: Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
| ngram_lm_scale | attention_scale | | ngram_lm_scale | attention_scale |
|----------------|-----------------| |----------------|-----------------|
| 0.3 | 1.5 | | 0.5 | 1.3 |
To reproduce the above result, use the following commands for training: To reproduce the above result, use the following commands for training:
``` ```
cd egs/gigaspeech/ASR/conformer_ctc cd egs/gigaspeech/ASR
./prepare.sh ./prepare.sh
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
./conformer_ctc/train.py \ ./conformer_ctc/train.py \
@ -31,12 +32,12 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
--lang-dir data/lang_bpe_500 --lang-dir data/lang_bpe_500
``` ```
and the following command for decoding and the following command for decoding:
``` ```
./conformer_ctc/decode.py \ ./conformer_ctc/decode.py \
--epoch 19 \ --epoch 18 \
--avg 8 \ --avg 6 \
--method attention-decoder \ --method attention-decoder \
--num-paths 1000 \ --num-paths 1000 \
--exp-dir conformer_ctc/exp_500 \ --exp-dir conformer_ctc/exp_500 \
@ -47,3 +48,29 @@ and the following command for decoding
The tensorboard log for training is available at The tensorboard log for training is available at
<https://tensorboard.dev/experiment/rz63cmJXSK2fV9GceJtZXQ/> <https://tensorboard.dev/experiment/rz63cmJXSK2fV9GceJtZXQ/>
Results using HLG decoding + whole lattice rescoring:
| | Dev | Test |
|-----|-------|-------|
| WER | 10.51 | 10.62 |
Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
| lm_scale |
|----------|
| 0.2 |
To reproduce the above result, use the training commands above, and the following command for decoding:
```
./conformer_ctc/decode.py \
--epoch 18 \
--avg 6 \
--method whole-lattice-rescoring \
--num-paths 1000 \
--exp-dir conformer_ctc/exp_500 \
--lang-dir data/lang_bpe_500 \
--max-duration 20 \
--num-workers 1
```
Note: the `whole-lattice-rescoring` method is about twice as fast as the `attention-decoder` method, with slightly worse WER.