Impaczus/motif
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Stabilize SNES model audio chain.

Browse Source
This commit is contained in:
b1rdmania
2025-12-20 12:58:44 +00:00
parent 25a34b3519
commit 6a94031508
1 changed files with 63 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/synthesis/TestModelSynthesisEngine.ts
+63 -59
View File
@@ -34,10 +34,6 @@ export class TestModelSynthesisEngine {
private scheduleInterval = 25;
private fadeTime = 0.05;
// SNES-ish “APU constraints”
private snesSampleRate = 32000;
private snesBitDepth = 12;
constructor(audioContext: AudioContext, model: SynthModel) {
this.audioContext = audioContext;
this.model = model;
@@ -113,34 +109,74 @@ export class TestModelSynthesisEngine {
private configureMasterChain(): void {
try { this.masterGain.disconnect(); } catch {}
// dry always
const dry = this.audioContext.createGain();
dry.gain.value = 1.0;
this.masterGain.connect(dry);
dry.connect(this.postGain);
// Always connect master -> post, but for SNES we insert a stable “color” chain
// (bandlimit + gentle saturation + limiter) before post.
if (this.model !== 'snes') {
const dry = this.audioContext.createGain();
dry.gain.value = 1.0;
this.masterGain.connect(dry);
dry.connect(this.postGain);
return;
}
if (this.model !== 'snes') return;
// Create a single color chain (stable, avoids per-note ScriptProcessor crackle).
const colorIn = this.audioContext.createGain();
const hp = this.audioContext.createBiquadFilter();
const lp = this.audioContext.createBiquadFilter();
const shaper = this.audioContext.createWaveShaper();
const limiter = this.audioContext.createDynamicsCompressor();
// SNES-ish echo (APU has echo buffer + 8-tap FIR). Approx with feedback delay + lowpass.
// Band-limit similar to “sample playback through DSP”
hp.type = 'highpass';
hp.frequency.value = 45;
hp.Q.value = 0.7;
lp.type = 'lowpass';
lp.frequency.value = 5200;
lp.Q.value = 0.7;
// Gentle saturation curve (keeps things from sounding like pure oscillators)
// Type cast avoids TS lib generic mismatch (runtime is correct).
shaper.curve = this.makeSoftClipCurve(0.75) as any;
shaper.oversample = '2x';
// Limiter to stop “mess” / clipping during dense passages
limiter.threshold.value = -14;
limiter.knee.value = 18;
limiter.ratio.value = 12;
limiter.attack.value = 0.003;
limiter.release.value = 0.12;
// Wire: master -> colorIn -> hp -> lp -> shaper -> limiter -> post
this.masterGain.connect(colorIn);
colorIn.connect(hp);
hp.connect(lp);
lp.connect(shaper);
shaper.connect(limiter);
limiter.connect(this.postGain);
// SNES-ish echo (single instance): feedback delay + lowpass in loop
const wet = this.audioContext.createGain();
const delay = this.audioContext.createDelay(1.0);
const feedback = this.audioContext.createGain();
const fbFilter = this.audioContext.createBiquadFilter();
wet.gain.value = 0.28;
wet.gain.value = 0.22;
delay.delayTime.value = 0.165;
feedback.gain.value = 0.32;
feedback.gain.value = 0.28;
fbFilter.type = 'lowpass';
fbFilter.frequency.value = 1800;
fbFilter.frequency.value = 2200;
fbFilter.Q.value = 0.7;
this.masterGain.connect(delay);
// Tap echo from color chain output (post-filter) into delay, then back to post.
lp.connect(delay);
delay.connect(wet);
wet.connect(this.postGain);
delay.connect(fbFilter);
fbFilter.connect(feedback);
feedback.connect(delay);
}
private filterAssignments(assignments: RoleAssignment[]): RoleAssignment[] {
@@ -315,10 +351,7 @@ export class TestModelSynthesisEngine {
}
osc.connect(envelope);
// Per-voice crunch for SNES-ish
const out: AudioNode = this.model === 'snes' ? this.makeSnesCrunchNode(envelope) : envelope;
out.connect(layer.filterNode);
envelope.connect(layer.filterNode);
const gainValue = velocity * 0.5;
const attackBase = this.model === 'pre8bit' ? 0.003 : this.model === 'snes' ? 0.01 : 0.005;
@@ -339,7 +372,6 @@ export class TestModelSynthesisEngine {
try {
osc.disconnect();
envelope.disconnect();
if (out !== envelope) out.disconnect();
} catch {}
this.activeVoiceCount = Math.max(0, this.activeVoiceCount - 1);
}, (duration + releaseTime + 0.1) * 1000);
@@ -366,8 +398,7 @@ export class TestModelSynthesisEngine {
}
osc.connect(envelope);
const out: AudioNode = this.model === 'snes' ? this.makeSnesCrunchNode(envelope) : envelope;
out.connect(layer.filterNode);
envelope.connect(layer.filterNode);
const gainValue = (velocity * 0.3) / Math.max(chordPitches.length * 0.5, 1);
const attackBase = this.model === 'snes' ? 0.01 : 0.005;
@@ -388,49 +419,22 @@ export class TestModelSynthesisEngine {
try {
osc.disconnect();
envelope.disconnect();
if (out !== envelope) out.disconnect();
} catch {}
this.activeVoiceCount = Math.max(0, this.activeVoiceCount - 1);
}, (duration + releaseTime + 0.1) * 1000);
}
}
/**
* SNES “crunch” approximation:
* - downsample to ~32kHz
* - quantize to ~12-bit
*
* Implemented via ScriptProcessorNode for broad compatibility (lab-only).
*/
private makeSnesCrunchNode(input: AudioNode): AudioNode {
const sp = this.audioContext.createScriptProcessor(1024, 1, 1);
const gain = this.audioContext.createGain();
gain.gain.value = 1.0;
const targetRate = this.snesSampleRate;
const ratio = this.audioContext.sampleRate / targetRate;
const step = Math.max(1, Math.round(ratio));
const levels = Math.pow(2, this.snesBitDepth);
let hold = 0;
let last = 0;
sp.onaudioprocess = (e) => {
const inp = e.inputBuffer.getChannelData(0);
const out = e.outputBuffer.getChannelData(0);
for (let i = 0; i < inp.length; i++) {
if (hold-- <= 0) {
hold = step;
const q = Math.max(-1, Math.min(1, inp[i]));
last = Math.round(q * (levels / 2)) / (levels / 2);
}
out[i] = last;
}
};
input.connect(gain);
gain.connect(sp);
return sp;
private makeSoftClipCurve(amount: number): Float32Array {
// amount: 0..1 (higher = more distortion)
const k = 1 + amount * 30;
const n = 2048;
const curve = new Float32Array(n);
for (let i = 0; i < n; i++) {
const x = (i * 2) / (n - 1) - 1;
curve[i] = Math.tanh(k * x) / Math.tanh(k);
}
return curve as unknown as Float32Array;
}
private fadeOutAllLayers(): void {