Downloads · 30 days
0
icefall-asr-tedlium3-zipformer
desh2608/icefall-asr-tedlium3-zipformer
icefall-asr-tedlium3-zipformer is a machine learning model from desh2608. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Access
Public
Updated Jun 29, 2023
Repo size
1.1 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt1.1 GB · 98%
At a glance
- License
- apache-2.0
- Access
- Public
- Created
- Jun 16, 2023
- Updated
- Jun 29, 2023
- SHA
- b882505d
Datasets
Model card
From the Hugging Face model README
TedLium3 Zipformer
rnnt_type=regular
The WERs are
| dev | test | comment | |
|---|---|---|---|
| greedy search | 6.74 | 6.16 | --epoch 50, --avg 22, --max-duration 500 |
| beam search (beam size 4) | 6.56 | 5.95 | --epoch 50, --avg 22, --max-duration 500 |
| modified beam search (beam size 4) | 6.54 | 6.00 | --epoch 50, --avg 22, --max-duration 500 |
| fast beam search (set as default) | 6.91 | 6.28 | --epoch 50, --avg 22, --max-duration 500 |
The training command for reproducing is given below:
export CUDA_VISIBLE_DEVICES="0,1,2,3"
./zipformer/train.py \
--use-fp16 true \
--world-size 4 \
--num-epochs 50 \
--start-epoch 0 \
--exp-dir zipformer/exp \
--max-duration 1000
The tensorboard training log can be found at https://tensorboard.dev/experiment/AKXbJha0S9aXyfmuvG4h5A/#scalars
The decoding command is:
epoch=50
avg=22
## greedy search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 500
## beam search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 500 \
--decoding-method beam_search \
--beam-size 4
## modified beam search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 500 \
--decoding-method modified_beam_search \
--beam-size 4
## fast beam search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir ./zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 1500 \
--decoding-method fast_beam_search \
--beam 4 \
--max-contexts 4 \
--max-states 8
rnnt_type=modified
Using the codes from this PR https://github.com/k2-fsa/icefall/pull/1125.
The WERs are
| dev | test | comment | |
|---|---|---|---|
| greedy search | 6.32 | 5.83 | --epoch 50, --avg 22, --max-duration 500 |
| modified beam search (beam size 4) | 6.16 | 5.79 | --epoch 50, --avg 22, --max-duration 500 |
| fast beam search (set as default) | 6.30 | 5.89 | --epoch 50, --avg 22, --max-duration 500 |
The training command for reproducing is given below:
export CUDA_VISIBLE_DEVICES="0,1,2,3"
./zipformer/train.py \
--use-fp16 true \
--world-size 4 \
--num-epochs 50 \
--start-epoch 0 \
--exp-dir zipformer/exp \
--max-duration 1000 \
--rnnt-type modified
The tensorboard training log can be found at https://tensorboard.dev/experiment/3d4bYmbJTGiWQQaW88CVEQ/#scalars
The decoding commands are same as above.
- License
- apache-2.0
- Languages
- en
- Created
- Jun 16, 2023
- Updated
- Jun 29, 2023