nx_voiceprint.nx

nx_voiceprint.nx -- VOICE-CLONE-001 rung 1: a sovereign SPEAKER-IDENTITY descriptor ("voiceprint") extracted from a voice sample, plus a distance metric that tells two voices apart. A voice is a SOURCE (glottal pulses -> pitch, F0) driving a FILTER (the vocal tract -> formants). A speaker's IDENTITY lives mostly in the FILTER: the same person at a different pitch is still that person. LPC captures the vocal-tract filter exactly; its REFLECTION coefficients (PARCOR) are a compact, bounded (|k|<1), distance-friendly description of that filter. The voiceprint = the average PARCOR vector over the VOICED frames of a sample (voicing gated by nx_pitch) + the mean F0. Because we average only voiced frames and identity rides in the PARCOR (not the pitch), the print is pitch-invariant: the SAME speaker reading at 120 Hz vs 125 Hz prints nearly identical, while a DIFFERENT vocal tract prints far away. That property is the floor under voice DUPLICATION: capture who a speaker IS, separably from what they happen to be saying or how high they happen to be speaking. Composes (no reinvention): nx_lpc_autocorr (R[0..p]) -> nx_lpc_levinson (PARCOR k in Q30) ; nx_pitch (voiced-frame gate + F0). Integer-only, patent-clean. Voiceprint buffer layout (caller-owned *i64, VP_WORDS words): vp[0] = number of voiced frames averaged vp[1] = mean F0 (Hz) vp[2 .. 2+ORDER) = average reflection coeffs k[1..ORDER] in Q30 license_tier: ORIGINAL module: nishi-core.audio.voiceprint depends: nishi-core.audio.pitch, nishi-core.voice.lpc capability: AUDIO_SPEAKER_IDENTITY