Default test files are generating repeating text / hallucinating

[greenman2]

I used the default settings on the Kaggle notebook.

https://huggingface.co/datasets/sanchit-gandhi/whisper-jax-test-files

[sanchit-gandhi]

Looks like a hallucination (known issue of the Whisper model) - worth trying with timestamps

[sanchit-gandhi]

We could implement a logits processor that stops repetition of repeated tokens (similar to what we have in PyTorch: https://github.com/huggingface/transformers/blob/04ab5605fbb4ef207b10bf2772d88c53fc242e83/src/transformers/generation/logits_process.py#L474)

This might be quite difficult to do in a JAX compiled way though without returning the value of each token generated

[rpatel15-hue]

See this issue can be improved by reducing the --compression_ratio_threshold from default 2.4. openai/whisper#192
How can I modify this with this library?