Open fireindark707 opened 2 months ago
I'm currently not aware of any research or official documentation from OpenAI (Whisper's creators) or Guillaume Klein (faster_whisper's creator) showing improvements with this approach. I think using previous real-time transcriptions as prompts could potentially introduce biases / errors if the previous transcription was inaccurate. If you have specific examples or URLs showing improvements with these kind of prompts in faster_whisper, I'd be very interested in seeing them.
I believe an effective way to improve overall stability is to use the previously generated transcription results as the next prompt.