Skip to content

ADR-v2-017 — Emotion / Tone-Aware Formatting

Status: Accepted (2026-07-02) · Wave D Context links: [[adr-v2-000-interaction-layer]], [[adr-v2-002-prosody-autoformat]] (pause→sentence), [[adr-011]]

Context

Prosody→sentence punctuation (ADR-v2-002) formats from timing only. The Wave D research (#6) proposes affect-driven formatting: detect paralinguistic tone (excitement, question intonation, emphasis) and reflect it — auto !, emphasis on stressed words — using emotion2vec / SER front-ends (arXiv 2511.08723). Commercial dictation formats words but never tone.

Decision

Add an opt-in tone-aware formatting pass: [affect] enabled, mode (conservative default = emphasis + ! only | expressive), min_confidence. The formatting itself is a pure function (text, affect_label, confidence) → text (fully testable); the SER model is lazy behind an [affect] extra and, when absent, yields a neutral label so the pass is a no-op. Composes after Prosody Ink on the dictation path. OFF by default.

Consequences

  • The value-carrying logic (label→formatting) is pure and testable with no model.
  • Conservative default (emphasis + !) avoids mis-formatting neutral speech.
  • Caveat: SER is speaker/culture-variable → high min_confidence and conservative mode by default; expressive mode is explicit opt-in. On-device only (ADR-011).