Cleanup of `BatchedInferencePipeline`

SYSTRAN / faster-whisper

Faster Whisper transcription with CTranslate2

MIT License

12.66k stars 1.06k forks source link

Cleanup of `BatchedInferencePipeline` #1135

Closed MahmoudAshraf97 closed 1 week ago

MahmoudAshraf97 commented 1 week ago

This PR aims to further refine BatchedInferencePipeline class to include more transcription options and other improvements

Summary:

Supported new options for batched transcriptions:
- temperature
- hotwords
- fixed initial prompt
moved generate_segment_batched function from WhisperModel to BatchedInferencePipeline
make the pipeline class stateless
clarified documentation about what arguments do not have an effect in batched transcription

carolinaxxxxx commented 1 week ago

@MahmoudAshraf97 i check git+https://github.com/MahmoudAshraf97/faster-whisper.git@add_hotwords_batched and get error:

".../faster_whisper/transcribe.py", in line 487, in transcribe if not self.model.is_multilingual and language != "en": AttributeError: 'WhisperModel' object has no attribute 'is_multilingual'

MahmoudAshraf97 commented 1 week ago

fixed

carolinaxxxxx commented 1 week ago

@MahmoudAshraf97 i got another error after update:

"...faster_whisper/transcribe.py", line 545, in transcribe
all_language_probs=all_language_probs,
UnboundLocalError: local variable 'all_language_probs' referenced before assignment

MahmoudAshraf97 commented 1 week ago

@MahmoudAshraf97 i got another error after update:

"...faster_whisper/transcribe.py", line 545, in transcribe
all_language_probs=all_language_probs,
UnboundLocalError: local variable 'all_language_probs' referenced before assignment

Should be fixed