nx_voicegen_census.nx source

1// nx_voicegen_census.nx -- THE HONEST GAP: our voice/sound synthesis vs HUMAN-REALISTIC SOTA (operator 2026-07-04: 2// "none are human realistic ... we have a major gap on what you are saying and what is reality"). Grounded in the 3// REAL SOTA corpus fetched over our own TLS (arXiv papers + GitHub code + papers-with-code leaderboard, NOT just 4// Wikipedia): knowledge/fetched/sota_*.raw + voicegen_*.raw. Bar = human-realistic neural audio: neural VOCODERS 5// (WaveNet/HiFi-GAN, MOS ~4.4 near-human) + neural TTS (Tacotron2/FastSpeech2/VITS) + neural codec LMs (EnCodec/ 6// VALL-E voice-clone) + bioacoustic syrinx models. verdict = xcd_verdict(us,SOTA) COMPUTED; sovereignty tagged 7// [FLOOR] (real, but NOT realism); AHEAD liar-killed; neg-control fires. This census's JOB is to be BRUTAL about 8// the realism gap. license_tier: ORIGINAL expect_exit: 0 9import "nx_cms_exceed.nx" 10 11func sw(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } 12func sn(v: i64) -> i64 { let bb: *u8=sys_mmap(28); var m: i64=v; if m<0{m=0} let t: *u8=sys_mmap(28); var k: i64=0; if m==0{t[0]=48 as u8;k=1} while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} var i: i64=0; while i<k{bb[i]=t[k-1-i];i=i+1} sys_write(1,bb,k); return 0 } 13func have(path: *u8) -> i64 { let fd: i64=sys_openat_rd(path); if fd<0 {return 0} sys_close(fd); return 1 } 14 15// tot: [0]BEHIND [1]PARITY [2]AHEAD [3]INVALID [4]AHEAD-FLOOR [5]AHEAD-CAP [6]UNGROUNDED 16func row(dom: *u8, cap: *u8, our_c: i64, bar_c: i64, n: i64, ev: *u8, ground: *u8, floor: i64, tot: *i64) -> i64 { 17 let v: i64 = xcd_verdict(our_c, bar_c) 18 let ok: i64 = xcd_referee_ok(v, our_c, n) 19 let g: i64 = have(ground) 20 sw(" ["); sw(dom); sw("] "); sw(cap); sw(" -> "); sw(xcd_vname(v)) 21 sw(" (us "); sn(our_c); sw("/"); sn(n); sw(" vs SOTA "); sn(bar_c); sw(") "); sw(ev) 22 if floor==1 { sw(" [FLOOR: sovereignty, NOT realism]") } 23 if ok==0 { sw(" <== INVALID (LIAR-KILL)"); tot[3]=tot[3]+1 } 24 if g==0 { sw(" [UNGROUNDED: source absent]"); tot[6]=tot[6]+1 } 25 sw("\n") 26 if v==XCD_BEHIND { tot[0]=tot[0]+1 } 27 if v==XCD_PARITY { tot[1]=tot[1]+1 } 28 if v==XCD_AHEAD { tot[2]=tot[2]+1; if floor==1 {tot[4]=tot[4]+1} else {tot[5]=tot[5]+1} } 29 return 0 30} 31 32func main() -> i64 { 33 let tot: *i64 = sys_mmap(128) as *i64 34 var z: i64=0; while z<7 {tot[z]=0; z=z+1} 35 let K: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/" as *u8 36 let HIFI: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/sota_hifigan_fulltext.raw" as *u8 37 let VITS: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/sota_vits_paper.raw" as *u8 38 let VALLE: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/sota_valle_paper.raw" as *u8 39 let FS2: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/sota_fastspeech2_paper.raw" as *u8 40 let ENC: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/sota_encodec_paper.raw" as *u8 41 let PWC: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/sota_pwc_tts.raw" as *u8 42 let MOS: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/voicegen_mos.raw" as *u8 43 let SYR: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/knowledge/fetched/voicegen_syrinx.raw" as *u8 44 let OURS: *u8 = "/mnt/c/Users/elder/nishi-core/nxc2/runtime/nx_tts_speak.nx" as *u8 45 46 sw("=== nx_voicegen_census -- OUR formant synthesis vs HUMAN-REALISTIC neural SOTA (grounded in real papers+code) ===\n"); 47 sw("bar = neural audio (HiFi-GAN/VITS/VALL-E, MOS ~4.4 near-human). verdict COMPUTED. [FLOOR]=sovereign, NOT realism. scale 2=full 1=partial 0=absent.\n"); 48 49 sw("\n-- A. WHAT WE HAVE (sovereign FLOOR -- real, but does NOT make it sound human) --\n"); 50 row("SOV ","fully sovereign: pure integer/f32 math, ZERO ML weights, runs on 1 core, deterministic/reproducible", 1,0,1, "[ev: nx_tts_seq/speak = formant source-filter, no model, no training data, ~250KB]", OURS, 1, tot) 51 row("PIPE","a working TEXT->speech PIPELINE exists (G2P + phoneme synth + WAV), end to end sovereign", 1,0,1, "[ev: nx_tts_speak G2P dict+rules -> phoneme sequencer -> WAV on /listen]", OURS, 1, tot) 52 53 sw("\n-- B. THE REALISM GAP (the operator's point -- what we SAY vs REALITY; brutally BEHIND) --\n"); 54 row("NAT ","NATURALNESS: SOTA neural TTS = MOS ~4.4 (near-human, ground-truth ~4.5); ours = robotic formant, and NEVER MOS-measured", 0,2,2, "{sota_pwc_tts leaderboard + hifigan MOS ~4.3-4.4 vs human 4.45; formant synth caps ~MOS 3 = the honest gap}", PWC, 0, tot) 55 row("VOC ","NEURAL VOCODER: SOTA generates RAW WAVEFORM from a learned distribution (HiFi-GAN GAN / WaveNet AR / diffusion, trained on 1000s hrs); ours = 3 formant resonators", 0,2,2, "{sota_hifigan_fulltext: GAN vocoder, multi-scale/period discriminators; ours has NO learned waveform model}", HIFI, 0, tot) 56 row("ACM ","LEARNED ACOUSTIC MODEL (text->mel): Tacotron2/FastSpeech2/VITS learn from data; ours = hand-tuned formant tables + rule G2P", 0,2,2, "{sota_fastspeech2_paper: learned duration/pitch/energy; ours = static tables}", FS2, 0, tot) 57 row("E2E ","END-TO-END neural (VITS: text->waveform one model, MOS ~4.43, adversarial+flow+VAE); ours = hand-built cascade", 0,2,2, "{sota_vits_paper: MOS comparable to ground truth}", VITS, 0, tot) 58 row("CLONE","VOICE CLONING from seconds (VALL-E via neural codec tokens / EnCodec); ours = ONE hand-tuned voice, no cloning", 0,2,2, "{sota_valle_paper: 3-sec enrollment zero-shot clone; sota_encodec neural codec}", VALLE, 0, tot) 59 row("PROS","PROSODY: learned pitch contour + phoneme duration + energy (natural rhythm/emphasis); ours = FLAT monotone, fixed durations", 0,2,2, "{sota_fastspeech2: variance adaptor; ours = every phoneme same length, no pitch movement}", FS2, 0, tot) 60 row("SPEC","SPECTRAL FINE STRUCTURE: full envelope + breathiness/jitter/shimmer/aspiration; ours = 3 poles + white noise", 0,2,2, "{neural vocoders model the full spectrum; 3 formants = intelligible-not-natural}", HIFI, 0, tot) 61 row("BIRD","BIRDCALL realism: real birds = a SYRINX (two independent sound sources + nonlinear source-filter); ours = resonators+noise (the dog sounds)", 0,2,2, "{voicegen_syrinx: dual-source nonlinear organ; ours single-source linear = why the dog sounds are unconvincing}", SYR, 0, tot) 62 row("EVAL","EVALUATION: SOTA measures MOS with human panels (+ MCD/PESQ objective); we have ZERO realism measurement", 0,2,2, "{voicegen_mos: the metric; we have never scored our output -> can't even claim a number}", MOS, 0, tot) 63 64 sw("\n-- C. PARTIAL (a real start, still far below) --\n"); 65 row("G2P ","grapheme-to-phoneme: SOTA = large dict + learned/neural G2P + phoneme durations; ours = ~8-word dict + crude letter rules", 1,2,2, "{nx_tts_speak G2P is a real start but tiny vs a full pronunciation model}", OURS, 0, tot) 66 67 sw("\n=== SUMMARY ===\n"); 68 sw("BEHIND = "); sn(tot[0]); sw(" (naturalness/MOS, neural-vocoder, acoustic-model, end-to-end, voice-clone, prosody, spectral, birdcall, evaluation)\n"); 69 sw("PARITY = "); sn(tot[1]); sw("\n"); 70 sw("AHEAD = "); sn(tot[2]); sw(" of which capability="); sn(tot[5]); sw(" FLOOR(sovereignty)="); sn(tot[4]); sw("\n"); 71 sw("INVALID(liar-killed) = "); sn(tot[3]); sw(" UNGROUNDED = "); sn(tot[6]); sw("\n"); 72 sw("TOTAL GRADED = "); sn(tot[0]+tot[1]+tot[2]); sw("\n"); 73 74 sw("\n-- NEG-CONTROL (must be CAUGHT) -- fabricated 'our formant voice is already human-realistic / MOS 4.4' (us 0/2) -> "); 75 let neg: i64 = xcd_referee_ok(XCD_AHEAD, 0, 2) 76 var neg_caught: i64 = 0 77 if neg==0 { sw("REJECTED (liar-kill fired)\n"); neg_caught=1 } else { sw("ACCEPTED (BROKEN)\n") } 78 79 sw("\n=== THE HONEST TRUTH + THE REAL PATH ===\n"); 80 sw(" TRUTH: formant synthesis is a DEAD END for realism (caps ~MOS 3, unmistakably robotic). 'Elara has a voice' was overclaimed -- she makes an intelligible robotic SOUND, not a human voice. Human-realism REQUIRES a NEURAL vocoder + neural acoustic model, learned from real speech. There is no formant tweak that closes this.\n"); 81 sw(" THE PATH (sovereign): the SAME f32/SSE substrate that now runs the FIXED LLM forward + the Z-Image DiT is what runs a NEURAL VOCODER. V1 load a pretrained HiFi-GAN (small ~14M params) on our engine = real neural waveform gen. V2 a neural acoustic model (VITS/FastSpeech2) text->mel. V3 EnCodec/VALL-E for voice-clone. V4 MOS eval harness so we MEASURE realism, never claim it. Birdcall: a dual-source syrinx model.\n"); 82 sw(" FIRST MOVE: build the MOS/objective EVAL harness so every future claim is a measured number, not an adjective -- then load a neural vocoder and MEASURE the jump.\n"); 83 84 sw("\n=== VERDICT ===\n"); 85 var green: i64 = 1 86 if tot[3] != 0 { green = 0 } 87 if neg_caught != 1 { green = 0 } 88 if green==1 { 89 sw("GREEN -- census honest: 0 fabricated realism, liar-kill armed, neg-control fired. TRUTH: we are AHEAD only on the FLOOR ("); sn(tot[4]); sw(" sovereignty) and BEHIND on ALL "); sn(tot[0]); sw(" realism axes. The operator is RIGHT: what I SAID ('she has a voice') vs REALITY (robotic formant sound) is a MAJOR gap. Human-realistic = NEURAL, on the substrate we already fixed. Grounded in real SOTA (arXiv+GitHub+papers-with-code), not Wikipedia.\n"); 90 sys_exit(0); return 0 91 } 92 sw("RED -- census integrity failed\n"); sys_exit(1); return 1 93}