Silero based audio chunker

[yujonglee]

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

https://github.com/fastrepl/hyprnote/blob/573d5ef602b93638e7dc73a87ed64917d18bfc8a/plugins/local-stt/src/server.rs#L148-L150

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

• Max 30sec(Whisper constraint). Users want to see result in faster tempo. Targeting around 12se. Might need some scoring mechanism. VAD_prob * buffer_length
• 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:

https://github.com/fastrepl/hyprnote/blob/573d5ef602b93638e7dc73a87ed64917d18bfc8a/crates/chunker/src/predictor.rs#L14-L25

Max-length constraint:

https://github.com/fastrepl/hyprnote/blob/573d5ef602b93638e7dc73a87ed64917d18bfc8a/crates/chunker/src/stream.rs#L70

---

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.

https://github.com/fastrepl/hyprnote/blob/573d5ef602b93638e7dc73a87ed64917d18bfc8a/crates/chunker/src/lib.rs#L33

[yujonglee]

Also we have simple Silero-based VAD inference code here(https://github.com/fastrepl/hyprnote/blob/main/crates/vad/src/lib.rs).

So https://github.com/emotechlab/silero-rs is optional.

[yujonglee]

#867 works on test, but not with the actual application.
Some problem with the Stream implementation? Should fix test first