Music Theory for AI

AI music systems learn implicit music theory from training data. Understanding the fundamentals helps explain why certain prompts work, what the model has likely internalized, and where current systems still struggle.

Pitch and the Chromatic Scale

Western music divides the octave into 12 semitones. The frequency relationship between semitone $n$ and a reference pitch $f_{\text{ref}}$ (usually A4 = 440 Hz) is:

$$f(n) = f_{\text{ref}} \cdot 2^{n/12}$$

This exponential relationship means that pitch is fundamentally a logarithmic dimension — which is why log-frequency representations (mel, CQT) are more natural for music than linear frequency.

Intervals

An interval is the distance between two pitches, measured in semitones:

Semitones	Name	Sound Character
0	Unison	Identity
1	Minor 2nd	Tense, dissonant
2	Major 2nd	Stepping motion
3	Minor 3rd	Sad, dark
4	Major 3rd	Bright, happy
5	Perfect 4th	Open, suspended
7	Perfect 5th	Strong, stable
12	Octave	Same note class

AI models encode interval relationships implicitly. When a model generates a chord progression, it is navigating learned distributions over interval patterns.

Scales and Keys

A scale is a subset of pitches from the chromatic set. The two most fundamental:

• Major scale: W-W-H-W-W-W-H (W = whole step, H = half step)
• Minor scale (natural): W-H-W-W-H-W-W

A key specifies both a root note and a scale, establishing a tonal center. Most popular music operates within a single key or moves between closely related keys.

Key Signatures in AI Context

AI music generators have learned strong priors for key consistency. Prompts that specify mood (happy, dark, melancholic) implicitly bias toward major or minor tonality. Models rarely generate truly atonal music unless the training data includes significant atonal material.

Chords and Harmony

A chord is three or more notes sounded simultaneously:

Chord Type	Intervals (semitones from root)	Feel
Major triad	0, 4, 7	Bright, resolved
Minor triad	0, 3, 7	Dark, emotional
Diminished	0, 3, 6	Tense, unstable
Augmented	0, 4, 8	Mysterious, unresolved
Dominant 7th	0, 4, 7, 10	Bluesy, wants to resolve
Major 7th	0, 4, 7, 11	Jazzy, smooth
Minor 7th	0, 3, 7, 10	Warm, mellow

Chord Progressions

Common progressions appear frequently in training data and are strongly encoded in models:

• I–V–vi–IV (Pop: C–G–Am–F) — the most common pop progression
• ii–V–I (Jazz standard cadence)
• I–vi–IV–V (Classic doo-wop / 50s progression)
• i–bVII–bVI–V (Andalusian cadence, flamenco/metal)
• I–IV–V–I (Blues/rock foundation)

Models with text-to-music training have absorbed these patterns deeply. Prompt cues like "jazz," "pop," or "blues" activate corresponding harmonic priors.

Rhythm and Meter

Time Signatures

• 4/4: four beats per measure — overwhelmingly dominant in popular music and AI training data
• 3/4: waltz time
• 6/8: compound duple, common in ballads
• 5/4, 7/8: odd meters, rare in AI music output due to training data bias

Rhythmic Subdivisions

$$\text{Beat duration} = \frac{60}{\text{BPM}} \;\text{seconds}$$

Subdivision	Relationship
Quarter note	1 beat
Eighth note	1/2 beat
Sixteenth note	1/4 beat
Triplet	1/3 of a beat

Groove and Swing

Swing shifts every other subdivision away from the grid:

$$t_{\text{swung}} = t_{\text{grid}} + \delta \cdot \text{swing\_ratio}$$

AI models learn groove from training data. Specifying "swung," "syncopated," or "straight" in prompts can influence rhythmic feel, though control precision varies.

Song Structure

Most Western popular music follows sectional forms:

Section	Typical Function
Intro	Establish mood, introduce elements
Verse	Tell the story, lower energy
Pre-chorus	Build tension toward chorus
Chorus	Hook, highest energy, main melody
Bridge	Contrast, new harmonic area
Drop	Peak energy (EDM-specific)
Breakdown	Stripped-back, tension builder
Outro	Wind down, resolve

AI models have learned these structural conventions. Explicit structure tags in prompts (e.g., intro → verse → chorus → bridge → chorus → outro) help guide the generation trajectory.

Dynamics and Expression

Dynamic markings indicate loudness levels:

Marking	Level
pp (pianissimo)	Very soft
p (piano)	Soft
mp (mezzo-piano)	Moderately soft
mf (mezzo-forte)	Moderately loud
f (forte)	Loud
ff (fortissimo)	Very loud

Crescendo (gradually louder) and decrescendo (gradually softer) create energy arcs. AI models learn dynamic curves from training data, with builds and drops being particularly well-represented in EDM-heavy datasets.

Timbre and Orchestration

Timbre — the "color" of a sound — is determined by:

1. Harmonic content: overtone distribution
2. Spectral envelope: amplitude shape across frequencies
3. Temporal envelope: ADSR (Attack, Decay, Sustain, Release)
4. Noise components: breath, bow noise, pick attack

AI models encode timbre in embedding space, allowing prompts to specify instruments and production characteristics. The specificity of timbre control depends on how well-represented an instrument is in training data.

Why This Matters for AI Music

Music Theory Concept	How AI Models Use It
Pitch / intervals	Encoded in spectral representations
Scales / keys	Implicit statistical prior from training
Chord progressions	Sequence patterns in latent trajectories
Rhythm / meter	Temporal structure in token sequences
Song structure	State transitions during generation
Dynamics	Energy envelope in latent space
Timbre	Embedding clusters for instruments

Understanding music theory helps you write better prompts, diagnose generation problems, and interpret model behavior.