ADR-v2-104 — Prosodic Auto-Punctuation (acoustic, word-free sentence structure)¶

Status: Accepted (2026-07-02) · Wave L Context links: voice-punctuation (spoken words), prosody (pause→¶ + emphasis), itn, [[adr-011]]

Context¶

Wave L research (#10) — Voice Punctuation requires the user to say "comma"/"period"; Prosody Ink does pause→¶ and emphasis→bold only. Full sentence/clause punctuation driven by the acoustic prosodic contour (pause duration, pitch declination/reset, terminal rise) — inserting . , ? without any spoken punctuation words — is a distinct, broadly useful capability, and especially relieves RSI/motor users of saying "comma… period…" every clause. Anchor: Cho et al., Leveraging Prosody for Punctuation Prediction of Spontaneous Speech (Interspeech 2022, UW) — duration + pitch + energy features beat pause-only punctuation on ASR output.

Decision¶

Add an opt-in Prosodic Auto-Punctuation: [prosodypunct] enabled=false, sentence_pause_ms=700, comma_pause_ms=250. Pure core in prosodypunct/punctuate.py: punctuate_from_prosody(tokens, pause_after_ms, pitch_reset, terminal_rise, ...) → punctuated text — a sentence boundary (long pause or pitch reset) takes ? on a terminal rise else .; a medium pause takes ,. The F0/energy extraction is the optional lazy extra; the decision logic is 100% pure. OFF by default.

Consequences¶

• Natural punctuation from how you speak, no spoken punctuation words.
• Pure feature-rule engine → fully testable with synthetic prosodic vectors.
• Distinct from Voice Punctuation (acoustic vs spoken keywords) and Prosody Ink (adds . , ? vs ¶).
• Privacy (ADR-011): local prosodic features only.
• Caveat: heuristic thresholds (a learned model is a later tier); off by default.