feat: add audio chord detection for spectrogram regions

This commit is contained in:
Xiaohan-Tian
2026-05-25 19:09:19 -07:00
parent a7158b1ee0
commit f12add935a
15 changed files with 931 additions and 8 deletions
+75
View File
@@ -0,0 +1,75 @@
import { describe, expect, it } from 'vitest';
import { detectChordsFromAudio, type AudioChordDetectionRequest } from './audioChordDetectionCore';
const SAMPLE_RATE = 44100;
function createSineChordPcm(frequencies: number[], durationSeconds: number): Float32Array {
const sampleCount = Math.floor(durationSeconds * SAMPLE_RATE);
const pcm = new Float32Array(sampleCount);
for (let sampleIndex = 0; sampleIndex < sampleCount; sampleIndex++) {
const time = sampleIndex / SAMPLE_RATE;
let sample = 0;
for (const frequency of frequencies) {
sample += Math.sin(2 * Math.PI * frequency * time);
}
pcm[sampleIndex] = (sample / Math.max(1, frequencies.length)) * 0.35;
}
return pcm;
}
function createRequest(windows: AudioChordDetectionRequest['windows'], pcm: Float32Array): AudioChordDetectionRequest {
return {
pcm,
sampleRate: SAMPLE_RATE,
clipStartOffsetSeconds: 0,
windows,
};
}
describe('audio chord detection', () => {
it('detects a major triad from synthetic audio', () => {
const pcm = createSineChordPcm([261.63, 329.63, 392.0], 2);
const [result] = detectChordsFromAudio(createRequest([
{ barIndex: 0, startBeat: 0, endBeat: 4, startSeconds: 0, endSeconds: 2 },
], pcm));
expect(result.symbol).toBe('C');
expect(result.confidence).toBeGreaterThan(0);
});
it('detects a minor triad from synthetic audio', () => {
const pcm = createSineChordPcm([220.0, 261.63, 329.63], 2);
const [result] = detectChordsFromAudio(createRequest([
{ barIndex: 0, startBeat: 0, endBeat: 4, startSeconds: 0, endSeconds: 2 },
], pcm));
expect(result.symbol).toBe('Am');
});
it('marks silent analysis windows as no chord', () => {
const pcm = new Float32Array(SAMPLE_RATE * 2);
const [result] = detectChordsFromAudio(createRequest([
{ barIndex: 0, startBeat: 0, endBeat: 4, startSeconds: 0, endSeconds: 2 },
], pcm));
expect(result.symbol).toBe('N');
expect(result.confidence).toBe(0);
});
it('keeps neighboring synthetic bars stable', () => {
const barA = createSineChordPcm([220.0, 261.63, 329.63], 2);
const barB = createSineChordPcm([220.0, 261.63, 329.63], 2);
const pcm = new Float32Array(barA.length + barB.length);
pcm.set(barA, 0);
pcm.set(barB, barA.length);
const results = detectChordsFromAudio(createRequest([
{ barIndex: 0, startBeat: 0, endBeat: 4, startSeconds: 0, endSeconds: 2 },
{ barIndex: 1, startBeat: 4, endBeat: 8, startSeconds: 2, endSeconds: 4 },
], pcm));
expect(results.map(result => result.symbol)).toEqual(['Am', 'Am']);
});
});
+55
View File
@@ -0,0 +1,55 @@
import { KGProject } from '../core/KGProject';
import { KGAudioRegion } from '../core/region/KGAudioRegion';
import { beatRangeToSeconds, getAudioRegionDisplayLengthBeats } from './globalTrackUtil';
export {
detectChordsFromAudio,
type AudioChordDetectionRequest,
type AudioChordWindow,
type DetectedAudioChord,
} from './audioChordDetectionCore';
import type { AudioChordWindow } from './audioChordDetectionCore';
export function buildAudioChordWindowsForRegion(
project: KGProject,
audioRegion: KGAudioRegion,
): AudioChordWindow[] {
const regionStartBeat = audioRegion.getStartFromBeat();
const visibleLengthBeats = getAudioRegionDisplayLengthBeats(project, audioRegion);
if (visibleLengthBeats <= 0) {
return [];
}
const regionEndBeat = regionStartBeat + visibleLengthBeats;
const beatsPerBar = project.getTimeSignature().numerator;
const startBarIndex = Math.floor(regionStartBeat / beatsPerBar);
const lastBeatExclusive = regionEndBeat - 1e-9;
const endBarIndexExclusive = Math.max(
startBarIndex + 1,
Math.ceil(Math.max(regionStartBeat, lastBeatExclusive) / beatsPerBar),
);
const windows: AudioChordWindow[] = [];
for (let barIndex = startBarIndex; barIndex < endBarIndexExclusive; barIndex++) {
const barStartBeat = barIndex * beatsPerBar;
const barEndBeat = barStartBeat + beatsPerBar;
const overlapStartBeat = Math.max(regionStartBeat, barStartBeat);
const overlapEndBeat = Math.min(regionEndBeat, barEndBeat);
if (overlapEndBeat <= overlapStartBeat) {
continue;
}
const startSeconds = audioRegion.getClipStartOffsetSeconds() + beatRangeToSeconds(project, regionStartBeat, overlapStartBeat);
const endSeconds = audioRegion.getClipStartOffsetSeconds() + beatRangeToSeconds(project, regionStartBeat, overlapEndBeat);
windows.push({
barIndex,
startBeat: overlapStartBeat,
endBeat: overlapEndBeat,
startSeconds,
endSeconds,
});
}
return windows;
}
+256
View File
@@ -0,0 +1,256 @@
import FFT from 'fft.js';
const FFT_SIZE = 4096;
const HOP_SIZE = 512;
const MIN_ANALYSIS_FREQUENCY = 55;
const MAX_ANALYSIS_FREQUENCY = 1800;
const ABSOLUTE_SILENCE_RMS = 0.0025;
const RELATIVE_SILENCE_RATIO = 0.2;
const MIN_WINDOW_DURATION_SECONDS = 0.08;
const ROOT_NAMES = ['C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#', 'A', 'A#', 'B'] as const;
export interface AudioChordWindow {
barIndex: number;
startBeat: number;
endBeat: number;
startSeconds: number;
endSeconds: number;
}
export interface AudioChordDetectionRequest {
pcm: Float32Array;
sampleRate: number;
clipStartOffsetSeconds: number;
windows: AudioChordWindow[];
}
export interface DetectedAudioChord {
barIndex: number;
startBeat: number;
endBeat: number;
symbol: string;
confidence: number;
rms: number;
}
export interface AudioChordDetectionProgress {
completedWindows: number;
totalWindows: number;
percent: number;
}
interface ScoredChord {
symbol: string;
score: number;
}
const HANN_WINDOW = (() => {
const window = new Float32Array(FFT_SIZE);
for (let i = 0; i < FFT_SIZE; i++) {
window[i] = 0.5 * (1 - Math.cos((2 * Math.PI * i) / (FFT_SIZE - 1)));
}
return window;
})();
function clamp(value: number, min: number, max: number): number {
return Math.max(min, Math.min(max, value));
}
function cloneWindow(window: AudioChordWindow): AudioChordWindow {
return { ...window };
}
function buildTriadCandidates(chroma: Float64Array): { best: ScoredChord; second: ScoredChord } {
let best: ScoredChord = { symbol: 'N', score: Number.NEGATIVE_INFINITY };
let second: ScoredChord = { symbol: 'N', score: Number.NEGATIVE_INFINITY };
for (let root = 0; root < ROOT_NAMES.length; root++) {
const rootEnergy = chroma[root];
const minorThird = chroma[(root + 3) % 12];
const majorThird = chroma[(root + 4) % 12];
const fifth = chroma[(root + 7) % 12];
const outsideEnergy = Math.max(0, 1 - (rootEnergy + minorThird + majorThird + fifth));
const majorScore = (rootEnergy * 1.2) + (majorThird * 1.0) + (fifth * 0.8) - (outsideEnergy * 0.35) - (minorThird * 0.5);
const minorScore = (rootEnergy * 1.2) + (minorThird * 1.0) + (fifth * 0.8) - (outsideEnergy * 0.35) - (majorThird * 0.5);
const candidates: ScoredChord[] = [
{ symbol: ROOT_NAMES[root], score: majorScore },
{ symbol: `${ROOT_NAMES[root]}m`, score: minorScore },
];
for (const candidate of candidates) {
if (candidate.score > best.score) {
second = best;
best = candidate;
} else if (candidate.score > second.score) {
second = candidate;
}
}
}
return { best, second };
}
function analyzeChordWindow(
pcm: Float32Array,
sampleRate: number,
startSeconds: number,
endSeconds: number,
): { symbol: string; confidence: number; rms: number } {
const startSample = Math.max(0, Math.floor(startSeconds * sampleRate));
const endSample = Math.min(pcm.length, Math.ceil(endSeconds * sampleRate));
const sampleCount = Math.max(0, endSample - startSample);
if (sampleCount === 0 || (endSeconds - startSeconds) < MIN_WINDOW_DURATION_SECONDS) {
return { symbol: 'N', confidence: 0, rms: 0 };
}
const input = new Float32Array(FFT_SIZE);
const fft = new FFT(FFT_SIZE);
const output = fft.createComplexArray() as number[];
const chroma = new Float64Array(12);
let totalEnergy = 0;
for (let i = startSample; i < endSample; i++) {
const value = pcm[i];
totalEnergy += value * value;
}
const rms = Math.sqrt(totalEnergy / sampleCount);
const totalHops = Math.max(1, Math.floor(Math.max(0, sampleCount - FFT_SIZE) / HOP_SIZE) + 1);
for (let hop = 0; hop < totalHops; hop++) {
input.fill(0);
const frameOffset = startSample + (hop * HOP_SIZE);
const available = Math.max(0, Math.min(FFT_SIZE, endSample - frameOffset));
let frameEnergy = 0;
for (let i = 0; i < available; i++) {
const weighted = pcm[frameOffset + i] * HANN_WINDOW[i];
input[i] = weighted;
frameEnergy += weighted * weighted;
}
if (frameEnergy < 1e-7) {
continue;
}
fft.realTransform(output, input as unknown as number[]);
fft.completeSpectrum(output);
const earlyFrameWeight = Math.max(0.25, 1.5 - (hop / totalHops));
for (let bin = 1; bin < FFT_SIZE / 2; bin++) {
const frequency = (bin * sampleRate) / FFT_SIZE;
if (frequency < MIN_ANALYSIS_FREQUENCY || frequency > MAX_ANALYSIS_FREQUENCY) {
continue;
}
const real = output[2 * bin];
const imaginary = output[(2 * bin) + 1];
const magnitude = Math.sqrt((real * real) + (imaginary * imaginary));
if (magnitude < 1e-6) {
continue;
}
const midiPitch = 69 + (12 * Math.log2(frequency / 440));
const roundedPitch = Math.round(midiPitch);
const pitchClass = ((roundedPitch % 12) + 12) % 12;
const centsFromPitchClass = Math.abs(midiPitch - roundedPitch);
const pitchWeight = Math.max(0, 1 - (centsFromPitchClass / 0.5));
const frequencyWeight = 1 / Math.max(frequency, 80);
chroma[pitchClass] += magnitude * magnitude * pitchWeight * frequencyWeight * earlyFrameWeight;
}
}
const chromaTotal = chroma.reduce((sum, value) => sum + value, 0);
if (chromaTotal <= 0) {
return { symbol: 'N', confidence: 0, rms };
}
for (let i = 0; i < chroma.length; i++) {
chroma[i] /= chromaTotal;
}
const { best, second } = buildTriadCandidates(chroma);
return {
symbol: best.symbol,
confidence: clamp(best.score - second.score + (best.score * 0.2), 0, 1),
rms,
};
}
function smoothDetectedChords(results: DetectedAudioChord[]): DetectedAudioChord[] {
if (results.length < 3) {
return results.map(result => ({ ...result }));
}
const smoothed = results.map(result => ({ ...result }));
for (let i = 1; i < smoothed.length - 1; i++) {
const previous = smoothed[i - 1];
const current = smoothed[i];
const next = smoothed[i + 1];
if (current.symbol === 'N') {
continue;
}
if (previous.symbol === next.symbol && previous.symbol !== 'N' && current.symbol !== previous.symbol) {
const surroundingConfidence = Math.max(previous.confidence, next.confidence);
if (current.confidence < surroundingConfidence * 0.85) {
current.symbol = previous.symbol;
current.confidence = Math.max(current.confidence, surroundingConfidence * 0.75);
}
}
}
return smoothed;
}
export function detectChordsFromAudio(
request: AudioChordDetectionRequest,
onProgress?: (progress: AudioChordDetectionProgress) => void,
): DetectedAudioChord[] {
const windows = request.windows.map(cloneWindow);
if (windows.length === 0) {
return [];
}
onProgress?.({
completedWindows: 0,
totalWindows: windows.length,
percent: 0,
});
const rawResults: DetectedAudioChord[] = [];
windows.forEach((window, index) => {
const analysis = analyzeChordWindow(
request.pcm,
request.sampleRate,
request.clipStartOffsetSeconds + (window.startSeconds - request.clipStartOffsetSeconds),
request.clipStartOffsetSeconds + (window.endSeconds - request.clipStartOffsetSeconds),
);
rawResults.push({
barIndex: window.barIndex,
startBeat: window.startBeat,
endBeat: window.endBeat,
symbol: analysis.symbol,
confidence: analysis.confidence,
rms: analysis.rms,
});
onProgress?.({
completedWindows: index + 1,
totalWindows: windows.length,
percent: Math.round(((index + 1) / windows.length) * 100),
});
});
const maxRms = rawResults.reduce((max, result) => Math.max(max, result.rms), 0);
const silenceThreshold = Math.max(ABSOLUTE_SILENCE_RMS, maxRms * RELATIVE_SILENCE_RATIO);
const filtered = rawResults.map(result => (
result.rms < silenceThreshold
? { ...result, symbol: 'N', confidence: 0 }
: result
));
return smoothDetectedChords(filtered);
}