Skip to content

ADR-v2-101 — Hesitation-Hold Endpointing

Status: Accepted (2026-07-02) · Wave L Context links: autostop (silence endpoint), self_repair (content), disfluency (filler removal), [[adr-011]]

Context

Wave L research (#7) — Hands-Free Auto-Stop ends the turn on silence; Mid-Utterance Self-Repair fixes content. Neither addresses the false-endpoint problem: when a user trails off with a filled hesitation ("annnd… uhh…", an elongated vowel, a nasal "mmm") rather than clean silence, the near-silence that follows is mis-read as "done" and the recognizer cuts them off mid-thought. Detecting the filled pause and extending the endpoint threshold is a distinct timing behavior that especially helps slow/aphasic/aging speakers. Anchor: Chatziagapi et al., Audio and ASR-based Filled Pause Detection (ACII 2022) — filled pauses hold the floor; silence after word-lengthening/filled pauses should be relabeled as pause, not sentence end.

Decision

Add an opt-in Hesitation-Hold Endpointing: [hesitation] enabled=false, commit_ms=800, hold_extra_ms=1200. Pure cores in hesitation/endpoint.py: is_filled_pause(f0_series, spectral_stability) → detects a flat-F0, spectrally-stable sustained vocalization, and endpoint_decision(silence_ms, last_was_filled_pause, commit_ms, hold_extra_ms) → hold | commit (the threshold is extended by hold_extra_ms right after a filled pause). The F0/spectral extraction is the lazy extra; the rule is pure. OFF by default.

Consequences

  • Fewer mid-thought cut-offs for people who think out loud; a turn-taking accessibility win.
  • Pure rule over pre-computed features → fully testable with synthetic streams.
  • Distinct from Auto-Stop (extends vs fixed silence) and Self-Repair (timing vs content).
  • Privacy (ADR-011): local features only.
  • Caveat: heuristic thresholds; a persistent hesitater could feel slower (tunable); off by default.