Milestone M9 · #95

Fart-Off DSP & Acoustic Analysis

Every entry in The Great Fart-Off ships in the binary under src/engines/<slug>.zig and is pinned by the regression guard in src/golden.zig. This page breaks down the measured acoustic telemetry and internal synthesis architectures across all twelve contestants.

Acoustic & spectral telemetry

From the src/golden.zig fingerprints on Baseline 1 ("kujamba karibu") and Shuzi Seed 42, each engine's default "house sound" separates into distinct acoustic regimes (zero_crossings acts as a direct proxy for spectral brightness/rasp vs. sub-bass weight, peak out of 32767 shows limiter ceiling drive, and rms_milli measures overall acoustic fatness):

Tier ratings & DSP breakdown

S-Tier: Biomechanical physical models

1. opus · Claude Opus 5.5 — 9.7 / 10 (Best Overall & Best Physical Model)

Opus wrote a full biomechanical valve simulation (v3) while keeping the original v2 engine bit-for-bit accessible behind --voice classic=1.

• Phrase-level declination (SylContext): Tracks where a syllable sits in its word and phrase (stressed, word_final, phrase_final, position), so the gas reservoir naturally depletes across a phrase and sputters out on the final syllable.
• Stick-slip valve & phonation threshold (p_th): The sphincter is a pulse-train valve that checks p_now / p_th on every flap. As pressure falls near threshold, cycles randomly misfire (fire_p). When wet is high, the valve sticks shut (span = 1.3 + 1.0 * d1[2]), piles up pressure, and releases with an extra-loud boost and a NoiseBurst splat.
• Minnaert bubbles: Spawns four concurrent Bubble voices whose sine frequency chirps upward as the bubble radius shrinks (self.freq += self.rise).
• Full phonetic timeline: Models prenasalized murmurs (mb, nd, ng), voiced-stop closure bars, place-of-articulation formant transitions, and a trousers low-pass filter (a_lp), all in bit-identical polynomial math.

2. sonnet · Claude Sonnet 5.5 — 9.5 / 10 (Best Texture, Wetness & Showcase)

sonnet produces the most ragged, wet, biologically unstable sound in the field (3,882 zero crossings on baseline1 and 13.38M RMS on shuzi seed 42).

• Dual detuned sphincter lips (Flapper): Models two separate sphincter lips (fa and fb, detuned by 1.0 + 0.017 + 0.012 * tension) beating against each other. Every cycle draws Gaussian period jitter, amplitude shimmer, odd-cycle period-doubling (sub), and tail misfires (drop_prob).
• Zero-delay-feedback (ZDF) SVFs: Implements TPT/ZDF state-variable filters (svfCoef with detTan) updated every 16 samples (CONTROL_PERIOD), plus a 6 dB/oct pre-emphasis differentiator (FORMANT_EMPHASIS) so F2/F3 vowel formants cut through clearly.
• Six-voice bubble pool & voiced "bwomp" thump: Spawns up to six upward-chirping bubbles and adds a saturated sub-bass pitch-dropping thump (40 + 80 * e_thump Hz) on voiced plosives (b, d, g).

3. fable · Claude Fable 5.1 — 9.3 / 10 (Best Glottal Pulse & Shared Reservoir)

fable replaces the additive sine wave with a Rosenberg glottal pulse (flap(ph, oq)) and separates the integrated flow (fl.flow, low-passed at 400 Hz for the fat low body) from the radiated flow derivative (fl.dflow, which creates the sharp acoustic snap when the sphincter closes at 62%/38% asymmetry).

• Shared phrase gas reservoir: A phrase shares one reservoir that drains across syllables inside a word and takes a partial breath between words.
• Prenasalized & trilled onsets: Parses Swahili prenasalized stops (mba, nde, ngo vs. velar ng') so they hum in on the nasal and burst on the release, and gives r a 26 Hz onset trill over the first 80 ms.
• Cephes deterministic math: Uses quadrant-reduced Cephes polynomials (sinCos, exp2F, log2F) and f64 phase accumulation for bit-identical renders across build modes.

A-Tier: Heavyweight contenders & specialists

4. sol61 · GPT-6.1 Sol — 8.9 / 10 (Best "Chesty Thunder" Baseline)

sol61 posts the highest baseline RMS in the competition (8,097,826), delivering room-filling low-end weight with no flags required.

• Power-shaped valve + alternating double-pulses: Shapes the valve opening as opening^(1.6 + 4.5 * tension) multiplied by an alternating subharmonic (1.0 - 0.48 * wet * (0.5 + 0.5 * sin(subphase))), creating a wet double-slap on every other cycle.
• 250 Hz control-stream splat trigger: Steps a seeded control stream at 250 Hz that fires smoothed turbulence jets (splat_smooth) and irregular flutter modulation.

5. argon4 · Gemini 4 Argon — 8.8 / 10 (Best Dynamic Formant Squelch & Plosive Pops)

argon4 models the terminal tract as a moving, wet cavity rather than a static filter bank.

• Wobbling SVF formants: Modulates the F1, F2, and low-tract SVF cutoffs at audio rate using both the mucosal flutter LFO and a 2.35x faster bubble_lfo (wet_mod), producing a squelchy vowel filter sweep as pressure drops.
• Dedicated plosive pop oscillator: Maps every consonant (including prenasalized releases mb, mp, nd, ng, nj) to a PlosiveSpec with its own resonant pop_freq_hz (110 Hz for b, 135 Hz for p, 260 Hz for k, 310 Hz for t) and upward pitch spike.

6. astra · GPT-6 Astra — 8.6 / 10 (Deepest Sub-Bass Rumble)

With only 360 zero crossings on baseline1 and 8.02M RMS, astra is the sub-bass specialist of the default baselines.

• Half-sine duty-cycle valve: Uses a tension-controlled open duty cycle (open_width = 0.48 - 0.25 * tension) where the valve follows sin^2(pi * pos / w) - 0.5 * w while open and seals shut otherwise.
• Per-cycle crackle spits: At each phase wrap (phase >= tau), rolls the PRNG for cycle_jitter, cycle_level, and a 58% chance to fire a decaying crackle cluster gated by the valve opening.

7. grok47 · Grok 4.7 — 8.5 / 10 (Best Slit Friction & Showcase Comedy)

Grok 4.7 pairs the funniest showcase phrase (--phrase "po po po") with a clever model of high-tension raspiness.

• Dual-inharmonic chatter & slit friction: Adds two inharmonic chatter oscillators (chatter_hz at 18–88 Hz and chatter_hz2 at ~2.5x). When tension closes the aperture (openness == 0), high-frequency dry noise forces through the closed slit (slit = tension * (1.0 - openness) * |c2|).
• Normalized 0..1 knobs: Uses 0..1 amounts rather than 1.0-centered multipliers (pressure=0.72, tension=0.55 default).

B-Tier: Punchy & musical, simpler source models

8. mistral4 · Mistral Large 4 — 8.1 / 10 (Best Brassy "Disappointed Trombone")

Expands the additive oscillator to 12 harmonics tilted by intPower(0.8 + 0.45 * tension, k) through constant-skirt-gain biquad formants (peak gain = Q) and an 8-burst exponential turbulence train, all using a 4096-entry comptime sine table for cross-mode bit-exactness. Reads more like a brassy bass-trombone raspberry than a chaotic biological valve.

9. gutphysics · MiMo Pro 2.6 (Codebuff) — 7.8 / 10 (Cleanest Classic Extension)

Absorbed from PR #99, gutphysics runs an all-f64 signal path, scales flutter rate by Berea physics (sqrt(pressure / tension)), passes the source through three parallel RBJ formant biquads (F1/F2/F3), and schedules five discrete turbulence grain pops per syllable.

10. swe2 · SWE-2 — 7.6 / 10 (Sub-Octave Specialist, Quiet Baseline)

Uses sums of incommensurate low-frequency sinusoids (gurgle_hz at ~7 & ~13 Hz, sput_hz at ~4, ~7.5 & ~11 Hz) for deterministic liquid gurgle AM/FM and sputter gating, plus a 7 ms closure dip before a 14 ms 62 Hz plosive thump. A * 0.62 master trim keeps its baseline polite (3.47M RMS), though its freestyle showcase puts 92% of its energy below 100 Hz.

11. kimi3 · Kimi K3 — 7.4 / 10 (Punchy Clipper, Minimalist Tract)

Features an effective 11 Hz one-pole stochastic sputter gate (clamp(lp_gate * 30.0 - 0.15, 0.0, 1.0)) and drives a pressure-dependent soft clipper into the 0.95 ceiling (31,129 peak across all golden tests), but models only a single formant frequency per vowel (vowelFormantHz) plus one low SVF boom filter applied post-envelope.

12. gemini38 · Gemini 3.8 Flash — 6.8 / 10 (Polite Baseline, Strong Showcase)

Combines six harmonics, a sub-flap, quadratic/cubic waveshaping, and random 900–1500 Hz wet bubble pops (wet_pop). However, multiplying turb_signal, wet_pop, and the final output by the decaying pressure envelope on top of the main amplitude envelope double-attenuates the default voice (peak = 18,975, rms_milli = 2,568,370), making its un-flagged baseline quieter than stock until showcase knobs (pressure=2.0, wetness=2.2, intensity=1.5) are applied.

Summary ballot (#95)

• Baseline winner: opus (or sonnet for a wetter, raspier default; sol61 for pure chest-thumping loudness).
• Showcase winner: sonnet (pressure=1.5,tension=0.7,boom=1.8,decay=1.2,wetness=1.8,rasp=1.4,turbulence=1.2), with fable and grok47 (po po po) as runners-up.