VOICE_V1 voice tuning acceptance
Status: Accepted for VOICE_V1.
This note records the integrated Phase 7 tuning decision after T-46 through T-50 landed. The comparison workflow is scripts/bb8-metrics <reference-recordings-dir> target/bb8-metrics, which decodes the local BB-8 reference clips, renders the fixed dootdoot corpus, and reports directional metrics. These numbers are not a golden contract; they are the tuning aid used before the Phase 8 hash freeze.
| 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 | 0.543 | 185.760 | 1977.338 | 3552.979 | 667.529 | 0.951 | 0.571 | 0.488 | 0.059 | 0.001 |
Acceptance rationale:
- body: the low-body layer now lands in the same broad region as the BB-8 reference corpus instead of the near-zero sub-500 Hz baseline from the pre-tuning voice.
- upper-mid brightness: the rolloff and 2-5 kHz band moved upward while keeping >6 kHz modest, matching the reference caveat that brightness mainly lives in the upper-mid region, not the high treble.
- gesture motion: internal pitch/vowel trajectories and compound warble give single tokens motion without adding runtime randomness.
- harmonicity: transient/body/sparkle layers reduce the perfectly clean oscillator identity while preserving a pitched formant core.
- phrase air: the envelope and pause template reduce active density and leave more space between phrase units.
Known residual tradeoff: dootdoot remains cleaner and more semantically regular than the local reference set, which contains many one-off sound-effect gestures. That is acceptable for v1 because the learnable four-axis mapping needs a consistent voice family rather than a catalog of unrelated effects.