feat: acoustic environment fingerprinting + local privacy risk scorer #3

New issue

Open

opened 2026-04-12 10:56:21 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-12 10:56:21 -07:00

Owner

Context

Linnet#20. Extends cf-voice#2 (YAMNet) with privacy-relevant acoustic signal detection and a local risk scorer.

Signals

Scene classification: indoor/outdoor, vehicle, public space
Extended YAMNet labels: birdsong, traffic, street crossing, rain, background voices
Accent/language identification (separate lightweight model — whisper-lid or similar)
Background speaker presence (distinct from primary speaker via diarizer)

Privacy risk scorer

Runs locally — raw signals never leave the device
Returns a PrivacyRiskLevel: low | medium | high
High: background voices detected + outdoor public space → conversation may be overheard
Output: PrivacySignal event type on the SSE stream (Linnet surfaces this in UI)

Privacy inversion note

The signals that detect privacy-compromising environments are themselves potentially privacy-leaking. The scorer must run fully local; only the derived risk level (low/medium/high) may be transmitted, never the raw acoustic features.

Dependencies

cf-voice#2 (YAMNet base), cf-voice#1 (diarization — background speaker presence)

Tracking

Linnet#20

## Context Linnet#20. Extends cf-voice#2 (YAMNet) with privacy-relevant acoustic signal detection and a local risk scorer. ## Signals - Scene classification: indoor/outdoor, vehicle, public space - Extended YAMNet labels: birdsong, traffic, street crossing, rain, background voices - Accent/language identification (separate lightweight model — whisper-lid or similar) - Background speaker presence (distinct from primary speaker via diarizer) ## Privacy risk scorer - Runs locally — raw signals never leave the device - Returns a `PrivacyRiskLevel`: `low | medium | high` - High: background voices detected + outdoor public space → conversation may be overheard - Output: `PrivacySignal` event type on the SSE stream (Linnet surfaces this in UI) ## Privacy inversion note The signals that detect privacy-compromising environments are themselves potentially privacy-leaking. The scorer must run fully local; only the derived risk level (`low/medium/high`) may be transmitted, never the raw acoustic features. ## Dependencies cf-voice#2 (YAMNet base), cf-voice#1 (diarization — background speaker presence) ## Tracking Linnet#20