ADR-v2-002 — Prosody Auto-Formatting (pauses → punctuation, stress → emphasis)¶

Status: Accepted (2026-07-02) · Wave A Context links: [[adr-v2-000-interaction-layer]], existing Prosody Ink work (postprocess/, [prosody] config)

Context¶

Every dictation path collapses the prosodic layer (pitch/pause/energy) to flat text, so users must say "new paragraph" / "bold" (voice punctuation, v1.4.0). The voice-HCI research (internal) shows prosody is recoverable on-device and is "the discarded signal." YazSes already has partial Prosody Ink wiring; this ADR finishes it as a coherent auto-formatting feature.

Decision¶

Infer formatting from prosody, opt-in and conservative: 1. Pause → punctuation/paragraph (high confidence, cheap): use Whisper word timestamps — long inter-word gaps insert sentence punctuation; very long gaps or breath insert a paragraph break. Threshold-driven, tunable. 2. Stress/pitch → emphasis (lower confidence): optional parselmouth-based pitch/energy analysis marks emphasized spans (e.g. markdown **bold**) — behind a separate sub-flag and the prosody extra, so the DSP dependency stays optional.

Config: extend [prosody] — autoformat=false, pause_sentence_ms, pause_paragraph_ms, emphasis=false. Applied on the dictation path after clean_text, before injection; must compose with existing voice-punctuation and continuation spacing.

Consequences¶

• + Clean prose from stream-of-consciousness speech; less spoken markup.
• + Pause→punctuation reuses data already in hand (timestamps) — nearly free.
• − Over-formatting is annoying → ship conservative defaults, opt-in, easy to disable.
• − Pitch→emphasis needs a DSP lib and is error-prone → separate flag, optional extra, default off; validate on real dictation before recommending.