Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ASR: Is there any checked and stable way for pretrain? #11813

Open
ican24 opened this issue Jan 10, 2025 · 1 comment
Open

ASR: Is there any checked and stable way for pretrain? #11813

ican24 opened this issue Jan 10, 2025 · 1 comment

Comments

@ican24
Copy link

ican24 commented Jan 10, 2025

Hello,
I am trying to develop ASR tools for 2 custom languages.
I used 3-4 times speech_to_text_ctc_bpe.py for pretrain, but all cases I did not receive reslult. My best results is WER:0.98.

  
python speech_to_text_ctc_bpe.py \
	+config-name=../conf/fastconformer/fast-conformer_ctc_bpe \
	model.train_ds.manifest_filepath=/nemoasr/train.5 \
	model.validation_ds.manifest_filepath=/nemoasr/test.5 \
	model.tokenizer.dir=/nemoasr/tokenizer/tokenizer_spe_unigram_v512 \
	model.tokenizer.type=bpe \
	trainer.devices=-1 \
	trainer.accelerator="gpu" \
	trainer.strategy="ddp" \
	trainer.max_epochs=100 \
	model.optim.name="adamw" \
	model.optim.lr=0.001 \
	model.optim.betas=[0.9,0.999] \
	model.optim.weight_decay=0.0001 \
	model.optim.sched.warmup_steps=2000 \

Meanwhile I can get sufficient resluts (WER:3.5-4.5) using script. which I generated from examples in open sources of Internet (Japanese, Vietnamese).

Actually I would not to know is there any checked way for pretrain?
Thank you advance!

@ican24
Copy link
Author

ican24 commented Jan 15, 2025

No one knows pretrain mechanism???
It is so-so strange

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant