Phrase Detection

What are Musical Phrases?

A phrase is a musical sentence—a complete musical thought, typically 4, 8, 16, or 32 bars long.

Songs are built from phrases:

• Intro: Opening section (8-16 bars)
• Verse: Main storytelling section (16 bars)
• Chorus: Catchy, repeated section (8 bars)
• Bridge: Contrasting section (8 bars)
• Outro: Ending section (8-16 bars)

Detecting Phrase Boundaries

Phrase boundaries often occur where there are changes in the music:

• Energy changes (drums drop out, new instruments enter)
• Structural changes (verse to chorus)
• Silence or pauses
• Drum fills

// Detect phrase boundaries using energy changes
function detectPhrases(audioBuffer, beatPositions) {
    const energy = calculateEnergyEnvelope(audioBuffer);
    const boundaries = [];
    
    // Look for energy drops (common at phrase ends)
    const windowSize = 4; // 4 beats
    
    for (let i = windowSize; i < beatPositions.length - windowSize; i++) {
        const beforeBeat = beatPositions[i - windowSize];
        const afterBeat = beatPositions[i];
        
        const energyBefore = getAverage(energy, beforeBeat, beatPositions[i]);
        const energyAfter = getAverage(energy, beatPositions[i], afterBeat);
        
        // Significant energy drop
        if (energyAfter < energyBefore * 0.7) {
            boundaries.push({
                beat: i,
                time: beatPositions[i],
                strength: energyBefore - energyAfter
            });
        }
    }
    
    return boundaries;
}

Structural Analysis

Detect the overall structure (intro, verse, chorus, etc.) by analyzing patterns:

// Analyze song structure
function analyzeStructure(audioBuffer, phrases) {
    const sections = [];
    
    // Extract features for each phrase
    const features = phrases.map(phrase => ({
        energy: calculateEnergy(phrase),
        spectralCentroid: calculateCentroid(phrase),
        chroma: calculateChroma(phrase)
    }));
    
    // Cluster similar phrases
    const clusters = clusterPhrases(features);
    
    // Label sections
    // First cluster = verse (most common)
    // Highest energy cluster = chorus
    // First section = intro
    // Last section = outro
    
    return sections;
}

// Simple clustering
function clusterPhrases(features) {
    // Use similarity metrics to group phrases
    // Similar energy + timbre = same section
}

// MixMind's phrase detection
class PhraseDetector {
    async analyze(audioBuffer, beats) {
        // Calculate multi-band energy
        const energy = {
            bass: this.calculateEnergy(audioBuffer, 20, 250),
            mid: this.calculateEnergy(audioBuffer, 250, 4000),
            high: this.calculateEnergy(audioBuffer, 4000, 20000),
            total: this.calculateEnergy(audioBuffer, 20, 20000)
        };
        
        // Detect phrase boundaries
        const boundaries = this.detectBoundaries(energy, beats);
        
        // Group into phrases (typically 8 or 16 beats)
        const phrases = this.groupIntoPhrases(boundaries, beats);
        
        // Analyze structure
        const structure = this.analyzeStructure(phrases);
        
        return {
            phrases: phrases,
            structure: structure, // intro, verse, chorus, etc.
            phraseLength: this.detectPhraseLength(phrases)
        };
    }
    
    // Suggest mix points
    suggestMixPoints(structure) {
        const suggestions = [];
        
        // Intro end = good entry point
        if (structure.intro) {
            suggestions.push({
                time: structure.intro.end,
                reason: "End of intro - good for bringing in next track"
            });
        }
        
        // Before chorus = high energy entry
        if (structure.verse) {
            suggestions.push({
                time: structure.verse[0].start,
                reason: "Start of verse - build up to chorus"
            });
        }
        
        return suggestions;
    }
}

// Used for:
// - Suggesting mix-in/out points
// - Phrase-aligned transitions
// - Structure visualization