Music102: Theory & Computational Models

• Music102 is an integrative framework that combines introductory music theory education with advanced D12-equivariant computational models.
• It implements a Transformer architecture reparameterized for D12-equivariance, achieving higher chord prediction accuracy with fewer parameters.
• The ComposeOn Academy platform offers interactive, rule-based pedagogical tools that enhance music composition through guided theory and algorithmic analysis.

Music102 refers both to the paradigm of introductory music theory and composition education at the university level, and to a recent line of computational models, tools, and pedagogical resources for algorithmic accompaniment, symbolic music analysis, and digital composition. In recent research, Music102 designates both a $D_{12}$-equivariant transformer architecture for symbolic chord progression accompaniment (Luo, 2024) and the integrative pedagogical application ComposeOn Academy (Pu et al., 21 Feb 2025), as well as encompassing relevant algorithmic primitives for music information retrieval (Lerch, 2022). The following sections survey the mathematical, architectural, educational, and comparative aspects of the Music102 framework.

1. Mathematical and Theoretical Foundations

Music102, as instantiated in recent transformer models, leverages the structural symmetries of Western tonal music, formalized via the dihedral group $D_{12}$. The $12$ pitch classes of equal temperament form the cyclic group $C_{12}$ under transposition: $\mathcal{T}_i: P \mapsto P + i \pmod{12}$ and reflection symmetry, namely $\mathcal{R}: P \mapsto -P \pmod{12}$, collectively generate $D_{12}$: $$D_{12} = \{ \mathcal{T}_i, \mathcal{T}_i \mathcal{R} \mid i=0, \ldots, 11 \}$$ Chords are encoded as binary vectors $c \in \{0,1\}^{12}$, and sequences of melodies/chords as matrices in $\mathbb{R}^{12 \times T}$. A mapping $D_{12}$0 is $D_{12}$1-equivariant if: $D_{12}$2 This enforces transposition and reflection invariance: any data- or model-transformation commutes with the musical group action. The permutation representation $D_{12}$3 acts by permuting the pitch coordinates.

2. Model Architectures and Algorithmic Primitives

2.1. $D_{12}$4-Equivariant Transformer (Music102)

Music102's architecture comprises a Transformer encoder–decoder backbone, in which every subcomponent—including linear layers, nonlinearities, positional encodings, self-attention, and layer normalization—is reparameterized to satisfy $D_{12}$5-equivariance. The 12-dimensional input is decomposed into irreducible representations via orthogonal basis matrices $D_{12}$6, structuring the model's internal computations into equivariant channels.

Each equivariant linear layer acts as a block-diagonal operator, with parameters tied across pitch-class dimensions. Nonlinearities and normalization operate in the permutation basis and are projected accordingly. Multi-head attention in Music102 computes scores invariant to group actions, yielding $D_{12}$7-equivariant outputs.

2.2. Traditional Signal Processing and Analysis

Music102’s ecosystem includes algorithmic primitives supporting pedagogical and research applications (Lerch, 2022). Functionalities span:

• Low-level spectral features: Spectral centroid, spectral flux, and MFCCs.
• Fundamental frequency ($D_{12}$8) estimation: Autocorrelation, YIN.
• Onset detection: Time-domain energy and spectral-difference methods.
• Chord and key detection: Chroma features, template matching, Krumhansl-Schmuckler profiles.
• Sequence alignment: Dynamic time warping (DTW).
• Sequence decoding: Viterbi algorithm for HMMs.

These primitives, available as reference implementations (libACA, pyACA, ACA-Code), enable both hands-on analysis and the construction of symbolic music inputs for downstream learning systems.

3. Data, Training Procedures, and Quantitative Evaluation

3.1. Dataset and Preprocessing

The principal dataset for Music102’s transformer is POP909—comprising 907 pop songs with aligned melody and chord progression annotations. Melodies and chord progressions are discretized at the half-beat level into $D_{12}$9 and $12$0 representations, respectively.

3.2. Loss Function and Metrics

Training employs a weighted binary cross-entropy loss, emphasizing chord-changing timesteps. Evaluation metrics include:

• Weighted binary cross-entropy (BCE)
• Cosine similarity between predicted/ground-truth chords
• Exact-match accuracy per timestep

3.3. Empirical Results

Model (Params)	Weighted BCE ↓	Cosine Sim ↑	Exact Acc ↑
Music101 (6.85M)	0.5807	0.6638	0.1141
Music102 (0.76M)	0.5652	0.6727	0.1783

Music102 achieves significantly lower loss and higher accuracy while using approximately one-eighth the parameter count of its non-equivariant predecessor. This demonstrates the efficacy of group-theoretic parameter tying for pitch-class based music tasks (Luo, 2024).

4. Educational Platforms and Pedagogy: ComposeOn Academy

ComposeOn Academy is an integrative music theory-based environment designed for users with varying levels of musical expertise (Pu et al., 21 Feb 2025). Its architecture comprises:

• Audio→MIDI transcription via neural pitch tracking
• Symbolic feature extraction (chord/scale detection, scale-degree mapping)
• Rule-based chord progression extension (database of 39 progressions, 16 rhythm patterns)
• Staff and piano roll visualization
• Tiered music theory explanations (Beginner/Intermediate/Advanced levels)

The ComposeOn workflow anchors composition in the user’s input melody, analyzes scale/chord content, recommends functional progressions via string similarity matching, generates harmonic/melodic realizations, and supports measure-level user edits. A Music Theory Mentor chatbot interfaces with these explanations at hyperlinked terms, facilitating guided exploration.

ComposeOn supports interactive, stepwise composition and real-time feedback, contrasting one-shot generative outputs from black-box models.

5. Comparative Pedagogy and Systematic Evaluation

Evaluation with 10 users of varying music backgrounds showed ComposeOn consistently outperforming generative AI baselines (Suno) in perceived structure, coherence, and user confidence to compose (all with $12$1):

• Music-theory correctness: ComposeOn $12$2, Suno $12$3
• Structural coherence: ComposeOn $12$4, Suno $12$5
• Confidence/willingness to continue: ComposeOn > Suno on all dimensions

Qualitative feedback cited ComposeOn’s superior phrase grouping, more logical development, and adherence to user’s melodic ideas. The rule-based approach, extensive theory explanations, and interactive interface contributed to better learning outcomes and greater user agency (Pu et al., 21 Feb 2025).

6. Curriculum Implications and Integrative Practices

Music102, as a didactic framework, articulates compositional concepts—melody, scales, functional progressions, rhythm, ornamentation, and voice-leading—into explicit weekly modules. ComposeOn’s analytics and tiered explanation system support formative assessment and individualized pedagogical progression. Assignments emphasize Roman numeral analysis, progression justification, rhythmic/ornamental creativity, and theoretical reflection, facilitating deep integration of analysis and creation.

Composer-facing tools in the Music102 paradigm exemplify the synthesis of theory-driven and data-driven approaches, bridging classical musicological abstractions and contemporary algorithmic pipelines.

7. Limitations and Future Research Directions

Current $12$6-equivariant models assume strict equal-tempered symmetry and predominantly homophonic chord realization; richer polyphonic representations and additional group actions (including timbre/rhythm symmetry groups) remain open research problems. ComposeOn’s rule-based system privileges transparency and explainability over neural generative diversity, but may be less suited for highly idiosyncratic or genre-specific musical grammars. Further scaling, hybridization with neural methods, and extension to broader instructional contexts constitute natural next steps.

Music102 establishes a rigorous foundation for algorithmic music analysis, accompaniment, and pedagogy that is invariant under fundamental musical transformations, demonstrating advantages in both symbolic accuracy and compositional insight (Lerch, 2022, Luo, 2024, Pu et al., 21 Feb 2025).