VOICE_V10 acceptance — bidirectional whistle & gesture vocabulary

Status: Accepted for VOICE_V10. The --version string reports dootdoot VOICE_V10. This note records the directional gesture-vocabulary deltas behind the freeze; the binding contract is the regenerated golden WAV hashes (dootdoot-core/tests/fixtures/golden_wav_hashes.tsv, header # VOICE_V10), not the numbers here.

Source

Phase 19 implements the recommendations of bb8-sound-vocabulary-taxonomy.md: a frame-by-frame, gesture-level comparison found dootdoot had the right gesture families but was rising-biased and register-shy — it never produced a falling whistle, its accents barely left the register, single gestures spanned ~1 octave where BB-8 spans 3–4, neutral gestures ran long, and it never crossed into a rough/noisy burst.

The six slices (each its own revision):

• T-102 — a signed whistle sweep: a negative amount now descends toward WHISTLE_FLOOR_HZ instead of only climbing.
• T-103 — the exclamation terminal flourish descends (a falling whistle), while the question flourish keeps rising; direction follows pitch_velocity.
• T-104 — the accent whistle engages harder (an engaged floor instead of a near-zero ramp) and earlier (CURVE_WHISTLE_START_FRACTION 0.45 → 0.30), with the gate isolating the one promoted accent from non-accent body syllables.
• T-105 — the promoted semantic accent swoops on the wider ACCENT_PITCH_SPAN_SEMITONES (26 st) toward BB-8's multi-octave gestures.
• T-106 — neutral (text-path) syllables pace shorter than the base so gestures read as quick blips; the hand-built / empty-chirp path is byte-identical.
• T-107 — a body accent in an agitated utterance (high arousal and negative valence) bursts its noise/breath roughness toward the noisy band, then recovers; non-accent and calm syllables keep the steady-state texture.

Method

scripts/sound_taxonomy.py segments renders into gesture events and classifies each into a droid sound type (see the research note for the full method). The profile below pools ten emotionally-varied renders mirroring the BB-8 clip contexts (the same set the research note used), before (VOICE_V9) and after (VOICE_V10):

uv run scripts/sound_taxonomy.py target/sound-taxonomy/ddv10-wav/*.wav

This is a directional tuning instrument, not the voice contract; by-ear review remains the acceptance gate.

Vocabulary deltas (10 renders)

axis	BB-8 target	dootdoot V9	dootdoot V10	read
contour rise / fall / flat	31 / 28 / 41%	40 / 13 / 48%	34 / 22 / 45%	the fall fraction nearly doubled toward BB-8
falling gestures (down chirp/swoop/whistle)	many	2	13	descending gestures now exist (incl. chirp-down 0 → 7, a falling whistle)
gesture duration med / p90 (ms)	87 / 157	134 / 180	110 / 163	neutral gestures shortened toward BB-8's blips
per-gesture range p90 (st)	44.9	13.6	13.0	accents swoop wider in capability; p90 little moved on this sample
dominant-peak reach p90 (Hz)	2756	711	754	whistle reaches higher, but still rarely carries the peak
tonality mixed / noisy	16 / 2%	14 / 0%	18 / 0%	roughness lifted; no agitated-enough accent in this sample to cross into noisy

Read & residual gap

The headline win is polarity: dootdoot is no longer rising-only. The exclamation flourish descends, downward chirps/swoops appear across the corpus, and the fall fraction moves from 13% to 22% (BB-8 ~28%). Neutral gestures also shortened (134 → 110 ms median) toward BB-8's quick blips.

Two gaps remain capability-present but under-realised on this sample, consistent with the research note: the whistle now reaches higher and engages harder, but still rarely carries the dominant spectral peak for enough frames to move the p90, and the agitation burst (T-107) needs a genuinely high-arousal, negative-valence accent to fire, which this phrase set only grazes. Both are engagement-depth tuning, not missing machinery, and are left for a future pass.

Determinism & contract

• All synthesis stays inside the fixed, deterministic, bounded droid parameter space (NFR-16); no unseeded randomness was added.
• The hand-built / empty-chirp / neutral-curve path is byte-identical to VOICE_V9 (T-106 shortening and T-107 roughness are gated on the explicit text/planner path; the whistle sign/floor changes are no-ops at zero amount).
• ACTIVE_VOICE is VOICE_V10; the golden WAV hashes were regenerated and the double-run determinism test passes.