Downloads · 30 days
0
icefall_asr_aishell_whisper
yuekai/icefall_asr_aishell_whisper
icefall_asr_aishell_whisper is a machine learning model from yuekai. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Access
Public
Updated Jun 12, 2024
Repo size
24.6 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt24.6 GB · 100%
At a glance
- Access
- Public
- Created
- Jan 15, 2024
- Updated
- Jun 12, 2024
- SHA
- 79dede82
Model card
From the Hugging Face model README
Results
Aishell training results (Fine-tuning Pretrained Models)
Whisper
fine-tuning results on Aishell test set on whisper medium, large-v2, large-v3
| test (greedy search, before fine-tuning) | test (beam=10, after fine-tuning) | comment | |
|---|---|---|---|
| medium | 7.23 | 3.27 | --epoch 10 --avg 4, ddp |
| large-v2 | 6.56 | 2.47 | --epoch 10 --avg 6, deepspeed zero stage1 |
| large-v3 | 6.06 | 2.84 | --epoch 5 --avg 3, deepspeed zero stage1 |
Command for training is:
./prepare.sh --stage 30 --stop_stage 30
#fine-tuning with deepspeed zero stage 1
torchrun --nproc-per-node 8 ./whisper/train.py \
--max-duration 200 \
--use-fp16 1 \
--exp-dir whisper/exp_large_v2 \
--model-name large-v2 \
--deepspeed \
--deepspeed_config ./whisper/ds_config_zero1.json
# fine-tuning with ddp
torchrun --nproc-per-node 8 ./whisper/train.py \
--max-duration 200 \
--use-fp16 1 \
--exp-dir whisper/exp_medium \
--base-lr 1e-5 \
--model-name medium
Command for decoding is:
python3 ./whisper/decode.py \
--exp-dir whisper/exp_large_v2 \
--model-name large-v2 \
--epoch 999 --avg 1 \
--beam-size 10 --max-duration 50
NOTE: To decode with original whisper models, you should pad the input features into 30 secs. Otherwise it may not ouput EOS token.
Pretrained models, training logs, decoding logs, tensorboard and decoding results are available at https://huggingface.co/yuekai/icefall_asr_aishell_whisper
- Created
- Jan 15, 2024
- Updated
- Jun 12, 2024