Skip to content

ADR-v2-042 — Personal Read-Back Voice

Status: Accepted (2026-07-02) · Wave F Context links: [[adr-v2-035-speaking-coach]], Read-Back Loop (spec-read-back-loop), [[adr-v2-018-voice-biometric-guard]] (voiceprint), [[adr-011]]

Context

The Wave F research (#6) proposes read-back proofreading in a clone of the user's own voice from a short enrollment — hearing your own voice makes error-catching natural. Anchors: F5-TTS (CC-BY-NC), OpenVoice V2 (permissive → default), XTTS-v2 (non-commercial). The existing Read-Back Loop uses generic Kokoro voices; this personalizes and reuses voiceprint/ enrollment.

Decision

Extend [tts] with clone_voice=false and clone_backend="openvoice". The pure core clone_ready(reference_seconds, min_seconds) decides whether enrollment has enough audio, and select_clone_backend(config) returns the backend, defaulting to the permissively-licensed OpenVoice V2 (F5/XTTS are non-commercial and only used on explicit opt-in). The clone model is lazy behind the existing tts extra path. The voice-clone reference embedding is biometric → stored ONLY in the encrypted corpus (same rule as Voice Guard), never plaintext. OFF by default.

Consequences

  • Personalizes read-back; reuses voiceprint enrollment + the tts path.
  • Pure readiness/backend selection → fully testable with no model.
  • Licensing honored: permissive OpenVoice V2 default; non-commercial backends explicit opt-in.
  • Privacy (ADR-011/012): clone embedding biometric, encrypted-corpus only, never off-device.
  • Caveat: voice cloning is sensitive → off by default, local-only, explicit enrollment.