VOICE_V12 noun/verb recognizability — acceptance

Accepted for VOICE_V12 on 2026-07-04, after a two-round by-ear evaluation (2026-07-03, recorded in voice-v12-spike-evaluation.md) and the directional instrument checks below. The surfaced version string is dootdoot VOICE_V12.

What changed

Content words carry a systematic two-pillar class signature (FR-114…FR-121):

• Layered co-onset class markers — noun = broadband click/pop splash (7 inharmonic partials + 620 Hz thud, 30 ms window, 8 ms attack ramp, mix 0.10), verb = up-swept dual-sine chirp (1400→3600 Hz and 2050→5150 Hz, 50 ms, 25% attack fraction, mix 0.09). Both start together with the tonal body and bloom with its 15 ms attack so the mark fuses into the word rather than reading as a separate pre-beat (the round-1 by-ear finding; round 2 confirmed the fix and the softened gains).
• Compound stem → class-resolution silhouette — noun settles (vowel rounds toward oo, pitch steps down, contour flattens, warble calms), verb pushes (brighter toward ee, rising floor). Single-token content words gain one derived resolution syllable; multi-token words shape their last subword; every content-word syllable takes the 0.62 compound duration scale so words read heavy without dragging the VOICE_V11 pace.
• Classes from a baked, pinned table under the conservative ambiguity policy the by-ear A/B locked: noun/verb-ambiguous coding lemmas (build, fix, run, update) stay unmarked, closed-class words are excluded, and unclassified words render exactly as VOICE_V11 — the empty chirp and all hand-built paths are byte-identical.

By-ear acceptance

Round 1: class identity tellable (bug vs run blind), pacing breathes, but the markers read as a separate percussive pre-beat → fused attack ramps + softer mixes. Round 2: "the mark does now fuse into the word and softening sounds great." The conservative ambiguity leg won the A/B (the dominant leg's consistently-wrong markings were not preferred).

Directional instrument checks

scripts/sound_taxonomy.py on the frozen minimal pair — the classes separate on gesture category, not just pitch:

render	label	contour	tonality	net
bug	blip	flat	tonal	−1.2 st
run	chirp-up	rise	tonal	+9.2 st

scripts/acoustics (gap-analysis metrics) on "verify the bug in the changelog" vs the BB-8 reference clip: harmonicity median 0.879 vs 0.853 (in-family), max internal gap 35 ms vs 58 ms, 2–5 kHz share median 0.066 — the markers add bright-band attack energy while the voice stays one droid. Against the VOICE_V11 baseline the pacing guardrail held during the spike evaluation: identical island structure and +7% duration on the marked sentence.

Contract

ACTIVE_VOICE is VOICE_V12; the golden WAV fixtures were regenerated byte-for-byte (DOOTDOOT_REGEN_GOLDEN=1). Only text-path fixtures containing baked content words moved (cat, dash, ellipsis, long); empty.wav and all unclassified-word fixtures are unchanged, pinning the no-class path. The class table is part of the voice contract: regenerating assets/dootdoot_pos_v1.doot with different classifications changes rendered samples and requires a version bump.