Skip to content

ADR-v2-100 — Whisper-Aware Mode

Status: Accepted (2026-07-02) · Wave L Context links: acoustic_profiles (noise profile), voicehealth (vocal strain), vad_calibrated, [[adr-011]]

Context

Wave L research (#5) — Vocal-Strain Guard protects the voice; Acoustic Profiles picks a noise profile; neither detects whisper phonation and switches recognition behavior. Whispered speech is acoustically distinct (no F0/voicing, flatter spectral tilt, elevated turbulence) and normal STT degrades sharply on it. Detecting the whisper and auto-adapting gain, VAD threshold, and the STT prompt keeps quiet/whispered dictation accurate in shared or private spaces. Anchor: Ito et al., Analysis and recognition of whispered speech (ICASSP 2005); wTIMIT; Quartered Spectral Envelope 1D-CNN classification of phonated vs whispered speech (arXiv 2408.13746, 2024) — whisper is separable by voicing ratio + spectral tilt.

Decision

Add an opt-in Whisper-Aware Mode: [whispermode] enabled=false, voicing_max=0.3, tilt_min=-1.0, gain_db=6.0, vad_scale=0.5. Pure cores in whispermode/detect.py: voicing_ratio(frame) (normalized autocorrelation peak), spectral_tilt(frame, fs) (log-spectrum slope), is_whispered(feats, voicing_max, tilt_min) (low voicing and flat tilt), and whisper_adaptation(...) → a gain/VAD/prompt adjustment dict. Pure numpy DSP; no model. OFF by default.

Consequences

  • Accurate quiet/whispered dictation for weak-voice users and private spaces.
  • Pure DSP feature ratios + threshold rule → fully testable on synthetic frames.
  • Distinct from Vocal-Strain Guard (detect+adapt whisper vs protect voice).
  • Privacy (ADR-011): local audio only.
  • Caveat: threshold-based detection (a learned classifier is a later tier); off by default.