mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-08 08:34:19 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update RESULTS.md

Browse Source 
Lower WERs reported
This commit is contained in:
jinzr 2023-06-27 20:23:33 +08:00
parent f2a0a10482
commit 11faddc830
1 changed files with 47 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
egs/swbd/ASR/RESULTS.md
View File

@ -1,6 +1,53 @@
## Results
### Switchboard BPE training results (Conformer-CTC)
#### 2023-06-27
The best WER, as of 2023-06-27, for the Switchboard is below
Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
| | eval2000 | rt03 |
|--------------------------------|------------|--------|
| `conformer_ctc` | 30.80 | 32.29 |
Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
##### eval2000
| ngram_lm_scale | attention_scale |
|----------------|-----------------|
| 0.9 | 1.1 |
##### rt03
| ngram_lm_scale | attention_scale |
|----------------|-----------------|
| 0.9 | 1.9 |
To reproduce the above result, use the following commands for training:
```bash
cd egs/swbd/ASR
./prepare.sh
export CUDA_VISIBLE_DEVICES="0,1"
./conformer_ctc/train.py \
--max-duration 120 \
--num-workers 8 \
--enable-musan False \
--world-size 2 \
--num-epochs 100
```
and the following command for decoding:
```bash
./conformer_ctc/decode.py \
--epoch 99 \
--avg 10 \
--max-duration 50
```
#### 2023-06-26
The best WER, as of 2023-06-26, for the Switchboard is below