Cocktail Filter — origin note¶

Written: 2026-06-14 · Owner: Mohsen Seyedkazemi Ardebili Shipped as: design/v2-cognitive-layer/02-cocktail-filter.md · Companion: hci-corpus.bib Tier: design/ — public. This is a historical origin note, not a live proposal: the feature it describes has shipped, off by default, and the implementation doc above is the current source of truth.

The idea¶

The pitch was the human cocktail-party effect applied to a dictation daemon: enrol a voiceprint once, and treat every other voice in the room — a roommate on a call, a TV, an open-plan office — as if it were silence. The job story was simple: hold the key in a noisy kitchen, and only the enrolled voice reaches the transcript.

The state-of-the-art survey behind it split the problem into three layers of very different maturity. A personal-VAD gate — a small (~130K-parameter) speaker-conditioned classifier that keeps or drops whole frames by speaker — was judged solved for on-device, CPU-real-time use. Target-voice suppression (attenuating an interferer while keeping the target speaker intelligible, e.g. VoiceFilter-Lite-class models) was judged partial: demonstrated, but not available as a pip-installable, permissively-licensed CPU artifact. Full source separation (recovering both speakers cleanly) was judged out of scope — real, but GPU-tier, not a laptop-CPU feature.

What shipped¶

The gate layer shipped as src/yazses/audio/personal_vad.py, reusing the enrollment and encrypted-storage infrastructure YazSes already had. It is off by default: live testing after the initial build showed the sub-second window a real-time gate has to work with makes ECAPA-class speaker embeddings unreliable enough to false-reject the enrolled user's own voice — the opposite of the intended effect. That result, and the reasoning behind keeping the feature off pending a better short-window embedding, is recorded in design/v2-cognitive-layer/02-cocktail-filter.md and in the codebase itself. The suppression layer (P2) was never built — no permissively-licensed CPU-real-time model was found, exactly as the original survey flagged as the open risk.