Comparison to other open-source models and alternative training scenarios
#1
by
abilmansplus
- opened
Thank you for a great model!
It seems to perform better on open-source Kazakh datasets compared to whisper-turbo out-of-the-box.
Here are the results of my tests:
| model | WER on CV21 Kazakh test-set | WER on FLEURS Kazakh test-set |
|---|---|---|
| whisper-base-kk | 23.8% | 19.3% |
| whisper-turbo (beam_size=5) | 41.9% | 24.8% |
I have a couple of questions.
Why did you choose to fine-tune the base model?
Have you tried fine-tuning the turbo model?
Thank you.
Thank you for the additional tests! I chose the base model for deployment on an edge device for real-time applications. I haven't fine-tuned the turbo model, but the training recipe should be the same.
akuzdeuov
changed discussion status to
closed