WO-3 T7: SIMD blend shape solver from raw audio energy features
hunt
6b82c45
Eight bandpass resonators (RBJ constant-skirt biquads, transposed direct
form II) evaluated in f32x4 across BANDS, not samples: the filters are
recursive so the sample axis cannot be vectorised, while eight bands see the
identical input sample. f32x4 rather than f32x8 because this machine reports
sse4_2 with no AVX/AVX2/FMA, where wide emulates 256-bit as two halves.
Spectral shape (loudness removed) is mapped onto mouth channels by a fixed
basis with inhibitory terms: rounding is energy at 1.2 kHz IN THE ABSENCE OF
energy at 2 kHz, and without the negative weights every vowel would both
round and smile a little and converge on one mush shape.
Documented as a heuristic, not a trained regressor. TONGUE_OUT is never
driven -- tongue position behind closed lips has no acoustic signature.
23 tests. Measured on release:
solve: 2.686 us/frame for 267 samples (8 bands, f32x4) -- budget 200 us
jawOpen: /a/ = 0.6038, /i/ = 0.1996
/u/: funnel 0.1428 pucker 0.1048 smile 0.0000
/i/: fun…
