BB-8 word-attack forensics

Prompt

VOICE_V4 made repeated subword phrases more continuous, but the phrase I am so excited I am so excited I am so excited I am so excited still produced sharp, regular word attacks. The reference clip bb8-clips/inquisitive-then-chatty.mp3 has a softer onset that reads more like wah or woh.

Method

The investigation used three complementary checks:

• ffmpeg spectrograms and waveform crops for the generated phrase and inquisitive-then-chatty.mp3.
• A frame-RMS onset detector to locate large energy rises.
• A focused word-boundary metric: RMS level and derivative RMS in the first 18 ms after a bridged word boundary, divided by the following 45-85 ms body window.

The existing scripts/bb8-metrics report was also rerun to keep the finding in context. It still showed dootdoot as cleaner and more harmonically regular than the reference set, so the onset problem was treated as the first local fix rather than the whole timbre gap.

Finding

The harsh word attack was not the explicit transient fixed in VOICE_V4. Word starts were already connected and skipped that transient. The sharpness came from a level and shape mismatch:

• VOICE_V3/VOICE_V4 render word pauses as quiet bridges.
• The next connected syllable reused the high subword connection floor.
• The first 18 ms of the word body therefore jumped above both the bridge and the later vowel body.
• The vowel trajectory started near the semantic vowel target, so the onset sounded like a synth block beginning, not a rounded vocal opening.

Measured on the repeated excited phrase, VOICE_V4 word starts had a median word-start/body level ratio of 2.714 and roughness ratio of 2.376.

Fix Direction

VOICE_V5 separates word-boundary connections from subword connections:

• Subword connections keep the high floor needed for same-word continuity.
• Word-boundary starts ramp from a low bridge-matched floor into the syllable body.
• Word-boundary vowels start from a rounded oo-leaning pre-shape, then open into the semantic vowel target over a bounded window.
• Upper-mid sparkle and archetype texture are damped during that opening, then bloom back in.

After the change, the same phrase measured 0.375 for word-start/body level ratio and 0.361 for roughness ratio. The first phrase start remains allowed to be sharper; the fix targets bridged word starts inside a phrase.

Future Tools

Useful next forensic upgrades:

• Add a small in-repo onset report that emits level ratio, derivative ratio, zero-crossing rate, and spectrogram thumbnails for named fixtures.
• Add mel-band or Bark-band onset flux so upper-mid harshness can be tracked without over-weighting inaudible high harmonics.
• Add LPC/formant-trace estimates for reference clips so vowel opening can be compared directly instead of inferred from waveform shape.