Get segments based on VAD

[Ar770]

I'm currently employing a fine-tuned model that was converted to CT2 format. However, the word-level timestamps and segments produced by this model preform very badly.

As an alternative, I'm considering using a forced-alignment model, which performs optimally with audio chunks of approximately 8 seconds in length.

I'm curious to know if, instead of obtaining the original Whisper segments, it might be possible to acquire segments based on VAD.
Since the "max_speech_duration_s" parameter is available, I'm wondering if there's a way to achieve this segmentation based on VAD.

[toanhuynhnguyen]

If you want more accuracy for word and sentence timestamps, you can use Batched Faster-Whisper, I have tested it.
Or you can use https://github.com/m-bain/whisperX but it will be slower than Faster Whisper.