We use audio chunker to do whisper inference in streaming manner.

Good chunker is important. At the high-level, it should be

• Max 30sec(Whisper constraint). Around 12sec or so.
• Should split based on slience, should strip slience as much as possible. (Filter out silences #662) Whisper tends to hallucinate a lot on empty audio.

Our current approach:

chunker/stream.rs works with pluggable predictor.

Currently we use very simple RMS-based predictor:

Max-length constraint:

silero-rs is well-tested implementation.(Blog post)

We are not using it because it is hard to force 30sec max constraint. (emotechlab/silero-rs#31)

https://github.com/emotechlab/silero-rs/blob/6e8637b9d06cac41bbfe47e9933289f16ecbf87f/src/lib.rs#L85-L99

https://github.com/emotechlab/silero-rs/blob/6e8637b9d06cac41bbfe47e9933289f16ecbf87f/src/lib.rs#L387-L399

We have dataset to test if chunker works well or not.