Skip to content

ADR-v2-015 — Real-time Noise-Suppression Front-End

Status: Accepted (2026-07-02) · Wave D Context links: [[adr-v2-000-interaction-layer]], [[adr-v2-012-accessibility-continuum]] (quiet speech), [[adr-011]]

Context

YazSes has only an energy-VAD gate and no learned speech enhancement, so noisy/echoey rooms and quiet speech trip the "Silent audio — discarding" failure mode. The Wave D research (#3) points to DeepFilterNet3 (112K–3.58M params, real-time low-latency full-band enhancement, existing live plugin) as a drop-in denoise/dereverb stage.

Decision

Add an opt-in denoise front-end between mic capture and STT: [denoise] enabled, backend (deepfilternet | none), strength (graded). The backend is a pure audio-in → audio-out transform, lazy-imported behind a [denoise] extra; when off or absent it is a passthrough (identity), so the pipeline is unchanged. Pairs with Whisper/Low-Effort Mode (ADR-v2-012) to make quiet speech reliably clear the VAD gate.

Consequences

  • Directly attacks the recurring low-mic-level / discard failure mode.
  • Passthrough-when-off guarantees zero impact on the default path.
  • Caveat: over-aggressive suppression clips soft consonants → graded strength + advise re-running yazses mic-level after enabling. On-device only (ADR-011).