feat: add automatic filler word removal from transcriptions

[pchalasani]

Summary

• Add automatic filtering of filler words (uh, um, hmm, etc.) from transcriptions
• Remove hallucination patterns like [AUDIO], (pause), <tag>...</tag>
• Inspired by VoiceInk's TranscriptionOutputFilter approach

Changes

• Add filter_transcription_output() function in audio_toolkit/text.rs
• Integrate filter into transcription pipeline (runs after custom word correction)
• Add regex crate dependency
• Add comprehensive unit tests

Filler words removed

uh, um, uhm, umm, uhh, uhhh, ah, eh, hmm, hm, mmm, mm, mh, ha, ehh

Test plan

• Unit tests pass (cargo test --lib text)
• Manual testing with Parakeet V3 model - filler words are filtered correctly

🤖 Generated with Claude Code

[pchalasani]

Additional fix: Stutter collapse

Added collapse_stutters() to handle model hallucination artifacts where the model gets stuck repeating short tokens:

Before: "w wh wh wh wh wh wh wh wh wh why"
After: "wh why"

Before: "I I I I think so so so so"
After: "I think so"

Collapses any 1-2 letter word that repeats 3+ times consecutively to a single instance. Two repetitions like "no no" are preserved.

[cjpais]

I skimmed this code and I think it's gonna get pulled in. This seems pretty reasonable and definitely should be a default. Thank you for implementing this.

[pchalasani]

I skimmed this code and I think it's gonna get pulled in. This seems pretty reasonable and definitely should be a default. Thank you for implementing this.

Thanks. Incidentally I'm noticing a lot of stuttering issues with Parakeet V3, i.e. lots of repeated 2/3-letter words, like 5+ times

[cjpais]

Yeah, I get them quite frequently as well. And it happens if I do kind of a micro stutter, which normally you wouldn't notice but it really exaggerates it.

[pchalasani]

Analysis: Why VoiceInk has fewer stutters

I compared Handy's audio preprocessing with VoiceInk (which uses the same Parakeet V3 model but has noticeably fewer stuttering artifacts).

Key difference: VAD threshold

App	VAD Threshold	Effect
Handy	0.3	Very sensitive - captures marginal/unclear audio
VoiceInk	0.7	Conservative - only passes clear speech

Location in Handy: managers/audio.rs:121

let silero = SileroVad::new(vad_path, 0.3)

Why this matters

A lower threshold (0.3) means the VAD flags audio as "speech" even with low confidence. This sends more ambiguous audio to Parakeet, which then struggles and produces stuttering artifacts like "wh wh wh wh why".

VoiceInk's 0.7 threshold is more selective - it only passes audio that's clearly speech, giving the model cleaner input.

Proposed fix

Increase VAD threshold from 0.3 to 0.5 or 0.6. Trade-off: might occasionally clip very quiet speech, but should significantly reduce stuttering.

Would be happy to submit a follow-up PR with this change if you'd like to test it.

[cjpais]

I think if we're gonna tune the vad threshold, we should add it as a slider in the debug menu so we can adjust it directly there and find a good default more broadly. I think I'm quite hesitant to edit the default since, on the whole, it seems to work relatively well for most people. And your post-processing step should clean up Parakeet's issues. I would rather have more audio than less audio, generally speaking.

[cjpais]

Okay, I removed the bracketing changes mainly because I don't see them typically and I'd rather keep the code slim until we run into the issues for now.