Skip to content

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.

setclipsactive fractionactive island median msmagnitude centroid hzmagnitude rolloff 85 hzdominant peak motion hzharmonicitysub-500 hz power500-2000 hz power2000-5000 hz powerover-6000 hz power
BB-8 reference320.467290.2492382.0334651.1721335.0590.8060.5270.3280.0360.003
dootdoot corpus60.543185.7601977.3383552.979667.5290.9510.5710.4880.0590.001

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.

Independent open-source droid acoustics. Not affiliated with Lucasfilm or Disney.