VOICE_V2 expressiveness acceptance

Status: Accepted for VOICE_V2.

VOICE_V2 freezes the deterministic phrase, affect, complexity, and archetype performance channels documented in reference/voice_v2_scope.md. The four semantic axes remain the learnable core; the new channels only shape bounded performance around those axes.

VOICE_V3 superseded V2 by smoothing connected phrase rendering, VOICE_V4 smoothed repeated connected onsets, VOICE_V5 smoothed word attacks, and VOICE_V6 is the active voice contract. V2 remains the frozen expressiveness checkpoint.

Checks

• dootdoot --version reports dootdoot VOICE_V2.
• scripts/bb8-metrics <reference-recordings-dir> target/bb8-metrics-v2 renders the short BB-8 reference corpus and the contextual clips under bb8-clips/.
• The contextual metrics report is written to target/bb8-metrics-v2/contextual/ after decoding clips into target/bb8-metrics-v2/contextual-wav/.
• cargo test -p dootdoot-core --test golden_wav_hashes verifies the committed golden WAV hashes for the frozen v2 corpus.

Short-clip directional metrics:

set	clips	active fraction	active island median ms	magnitude centroid hz	magnitude rolloff 85 hz	dominant peak motion hz	harmonicity	sub-500 hz power	500-2000 hz power	2000-5000 hz power	over-6000 hz power
BB-8 reference	32	0.467	290.249	2382.033	4651.172	1335.059	0.806	0.527	0.328	0.036	0.003
dootdoot corpus	6	0.548	116.100	2139.833	3660.645	667.529	0.947	0.616	0.506	0.075	0.001

Contextual-clip directional metrics:

set	clips	active fraction	active island median ms	magnitude centroid hz	magnitude rolloff 85 hz	dominant peak motion hz	harmonicity	sub-500 hz power	500-2000 hz power	2000-5000 hz power	over-6000 hz power
BB-8 reference	7	0.988	499.229	1223.395	1744.189	2648.584	0.800	0.333	0.534	0.003	0.000
dootdoot corpus	6	0.548	116.100	2139.833	3660.645	667.529	0.947	0.616	0.506	0.075	0.001

Contextual Clips

The directional contextual clip set is:

• lost-friends-sad;
• excited-explanation;
• found-fixed-excitement;
• left-behind-anxious;
• explosion-surprise;
• enemy-approaching-alarm;
• inquisitive-then-chatty.

These clips are directional acceptance aids, not asset-spec constants. They check that sad, excited, anxious, surprised, alarm-like, and inquisitive phrases move the v2 performance channels in the intended directions while the golden WAV hashes remain the sample-level contract.

Decision

By-ear acceptance for the local v2 corpus is recorded as accepted. Phrase timing now creates deterministic boundaries, resets, lengthening, and declination. Affect supplies bounded valence/arousal mood shaping. Complexity adds bounded compound articulation from owned WordPiece and character signals. Archetype selection and rendering add fixed chatter, yelp, moan, stutter/burst, and tremble gesture families plus sparse deterministic servo/noise seasoning.