nx_phoneme_synth.nx

nx_phoneme_synth.nx -- VOICE-CLONE-001 rung 5a: render a PHONEME SEQUENCE into time-varying speech. This is the mechanism that turns a string of sounds into a spoken WORD -- the difference between a sustained vowel and reading a book. Source-filter, integer-only. Each phoneme is a DATA record (so the phoneme inventory lives in a table, not in code -- "data not code"): record = [ a1_F1, a1_F2, a1_F3, voiced(1/0), dur_samples ] (PH_REC words) a1_Fk = the 2-pole resonator coefficient 2*r*cos(2*pi*Fk/fs) in Q15 for formant k (precomputed per phoneme); shared pole radius r=0.95. voiced = 1 -> glottal impulse train at f0 ; 0 -> noise (fricatives etc.) The three formant resonators run in CASCADE and their state is CARRIED ACROSS phoneme boundaries -- so when the formant targets change, the filter rings smoothly into the new shape. That is coarticulation-lite: natural glides between sounds instead of clicks, which is what makes a sequence sound spoken. license_tier: ORIGINAL module: nishi-core.audio.phoneme_synth depends: (none -- pure integer DSP; mmap scratch only) capability: AUDIO_ARTICULATION