search

FL33TW00D / whisper-turbo

Cross-Platform, GPU Accelerated Whisper 🏎️
https://whisper-turbo.com
Apache License 2.0
1.72k stars 75 forks source link

[Feature request] Condition on previous text #64

Closed Huguet57 closed 7 months ago

Huguet57 commented 9 months ago

Hi, It's me again. It would be pretty cool to add the condition_on_previous_text flag to whisper-webgpu as in:

https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/transcribe.py#L278

Setting it to "false" can greatly reduce the failure loop cases. In minority languages is pretty usual.

Huguet57 commented 9 months ago

Basically what the flag does is deleting the prompt that passes through the windows.

FL33TW00D commented 9 months ago

@Huguet57 Unfortunately, we won't be adding any more features to whisper-turbo until ratchet reaches feature parity.

We are pretty close to running Whisper on the new OSS framework (see PR here).

Once we've done that, we intend to add all sampling methods from the original repo.

Huguet57 commented 9 months ago

Works for me! Even better if OSS. If any help testing is needed, let me know

FL33TW00D commented 7 months ago

Should be a 2 second job here: https://github.com/FL33TW00D/ratchet/blob/master/crates/ratchet-models/src/whisper/transcribe.rs#L97