MusicGen-Chord: Explicit Chord-Controlled Generation

• MusicGen-Chord systems integrate explicit, multi-hot chord conditioning into neural networks to produce harmonically coherent and diverse musical outputs.
• Architectural designs include Transformer-based audio generators and Retrieval-Edit-Rerank symbolic pipelines that enable fine-grained chord injection and modular control.
• Evaluation using metrics like chord entropy, coverage, and pitch–chord compatibility validates improved performance across both audio and symbolic music generation tasks.

MusicGen-Chord denotes a family of systems leveraging explicit chord progression features, advanced conditioning mechanisms, and principled architectural decomposition to generate music that is both harmonically coherent and musically diverse. The term encompasses both symbolic and audio-level generation approaches, with significant variation in methodologies, interpretability, and integration into real-world applications. The frameworks discussed here include end-to-end autoregressive audio models (as in MusicGen-Chord (Jung et al., 2024)), system-level decompositions for symbolic chord sequence generation (Retrieval-Edit-Rerank, or RER (He et al., 8 May 2026)), and integration of chord synthesis into remixing and online music creation workflows.

1. Core Principles and Definition

MusicGen-Chord denotes systems that inject explicit chord progression control into neural music generation pipelines. Unlike melody-only or text-only conditioning, these systems represent chordal information—typically as multi-hot chroma vectors, symbolic chord tokens, or explicitly enumerated progressions—which are processed alongside other conditioning signals to modulate the generative model at respective time scales.

Fundamental to MusicGen-Chord are:

• The use of explicit chord conditioning, often via dense chroma encodings applied per time frame rather than single-pitch or scalar labels
• Integration of symbolic chord progressions, either enumerated exhaustively via music theory grammars or retrieved from large existing corpora
• Decomposed or modular generative pipelines, which separate stylistic, music-theoretic, and preference-based considerations
• Facilitation of controllable and interpretable music generation, enabling both back-end API-driven workflows and interactive GUI-based composition

2. Architectural Designs

MusicGen-Chord architectures fall into two main categories: Transformer-based autoregressive generators with chroma or token-based conditioning, and symbolic sequence generators utilizing RNNs, HMMs, or system-level decomposition.

2.1 Transformer-based Audio Music Generation

The prototypical instantiation (Jung et al., 2024) modifies the original MusicGen (Transformer decoder over EnCodec tokens) by replacing the original 12-dimensional one-hot melodic chroma features with multi-hot chord chroma vectors. The time-indexed chroma matrices are parsed from symbolic chord progressions (using Harte encoding) or extracted from audio with chord recognition models such as BTC. These vectors are linearly projected and supplied to each layer via cross-attention:

$$C^{\mathrm{chord}}_t \in \{0,1\}^{12}, \quad \sum_{i=0}^{11} C^{\mathrm{chord}}_{t,i} \ge 1$$

No fine-tuning is applied; MusicGen's pretrained weights are reused, and the conditioning mechanism generalizes to the multi-hot case, enabling chord progressions to control the generator at inference time.

2.2 Retrieval-Edit-Rerank Symbolic Pipeline

In the RER paradigm (He et al., 8 May 2026), chord generation is decomposed into distinct modules:

1. Retrieval: Input melody is encoded into a 256-dim space via contrastive learning. FAISS-based nearest-neighbor search retrieves K=100 stylistically similar candidate chord sequences.
2. Editing: Candidates are projected into a feasible set by enforcing vertical (tonal alignment), horizontal (cadential resolution), and global (regularization) constraints. The cost function is minimized over the sequence using the Viterbi algorithm:

$$C_{e} = \underset{C \in \mathcal{F}}{\operatorname{argmin}}\, d(C, C_{r})$$

where

$$d(C, C_{r}) = \sum_{i} w^{(\mathrm{tonal})} d^{(\mathrm{tonal})}_i + w^{(\mathrm{cad})} d^{(\mathrm{cad})}_i + w^{(\mathrm{glob})} d^{(\mathrm{glob})}_i$$

1. Reranking: Feasible candidates are scored by a weighted sum of retrieval similarity and edit cost and the best candidate is returned.

This modular design allows fine control over diversity and feasibility, outperforming end-to-end models on chord entropy, coverage, and harmonicity metrics.

3. Input Encoding and Conditioning Mechanisms

Accurate rendering of chord progressions in both symbolic and audio domains relies on high-resolution, musically meaningful encoding:

• Chroma Matrices: Time-aligned 12-dim binary vectors, representing all pitch classes present in each chord
• Symbolic Tokens: Discrete chord types (e.g., 48 categories = 12 roots × 4 qualities), mapped per frame/bar
• Enumeration Grammars: Declarative music-theory-based libraries producing exhaustive sets of possible progressions for a fixed length and key, enabling integration with neural or retrieval-based generators (Lakshminarasimhan, 2024)
• Integration into Decoders: Cross-attention layers inject the chordal features into every decoder block, enabling persistent chord awareness during autoregressive audio token prediction

Chord progression encoding and injection mechanisms are explicit, not emergent, thus supporting direct manipulation and control by both human users and algorithmic agents.

4. Evaluation Methodologies

Systems are evaluated using both objective and subjective protocols:

4.1 Objective Metrics

Across domains, the following metrics are predominant:

• CHE (Chord Entropy Difference): Closeness of generated chord distribution entropy to ground truth
• CC (Chord Coverage Difference): Coverage over unique chord types compared to real music
• CTD (Chord Transition Distribution): Earth Mover's distance between real and generated transition matrices
• PCS (Pitch–Chord Compatibility Score): Proportion of melody notes matching active chord pitch classes
• MCTD (Mean Chord-Tone Distance): Average semitone distance between melody and chord tones
• CTnCTR (Chord-Tone/non-Chord-Tone Ratio): Fraction of chord tones relative to the melody

Subjective listening studies employ Likert scales for harmonicity, creativity, and overall preference, often divided by listener expertise.

4.2 Ablation and Analysis

Ablations demonstrate the necessity of each module:

• Removing retrieval collapses diversity (ΔCHE and ΔCC drop)
• Removing editing induces unmusical or invalid chord transitions (PCS falls, MCTD rises)
• Omitting reranking has smaller but consistent preference impact

Comprehensive analysis suggests that explicit retrieval injects human-like diversity, while editing and reranking safeguard feasibility and user preferences (He et al., 8 May 2026).

5. Applied and Interactive Systems

MusicGen-Chord frameworks are deployed in full-stack, web-accessible applications, facilitating both research and creative practice (Jung et al., 2024):

• MusicGen-Remixer: Accepts uploaded audio, infers chord progression (BTC), aligns detected chords to beats (All-in-One), generates new instrumental accompaniment conditioned on user-specified text and chords, and time-warps output for seamless remixing. Integration with Docker/cog enables reproducible, easily deployable endpoints.
• Mobile and API-based Chord Grammar Utilities: Exhaustively enumerate all music-theory-valid 4- and 8-chord progressions and expose REST and WebSocket APIs for use in DAWs, music theory education, or real-time interactive composition (Lakshminarasimhan, 2024)
• Symbolic Transformers and GANs: Autoregressive generation in the MMT-BERT framework with explicit chord token augmentation achieves state-of-the-art pitch-class entropy similarity, scale consistency, and groove preservation (Zhu et al., 2024)

These systems enable both granular control over harmonic content and reliable integration into live and batch music generation pipelines.

6. Limitations, Open Problems, and Future Directions

Current frameworks, while advancing control and interpretability, exhibit specific challenges:

• Editing may over-correct when candidate retrieval is far from the feasible set, yielding harmonically conservative outputs in rare but non-negligible cases (He et al., 8 May 2026)
• Chord conditioning typically operates at the level of pitch-class sets or discrete tokens, not with explicit voicings, inversions, or inner voice leading—limiting expressivity
• Integration of richer symbolic constraints (e.g., voice-leading, functional harmony) and dynamic λ-based reranking mechanisms is an open avenue
• Subjective evaluation often lags technical development; robust perceptual metrics and large-scale listening studies are required

Proposed directions include dynamic, learned controllers for diversity–fidelity trade-offs, multiresolution chord modeling, and joint generation of melody, chord, and style using unified multi-modal conditioning streams.

---

In summary, MusicGen-Chord systems embody explicit, interpretable conditioning on chord progressions for controllable, musically coherent generation, spanning symbolic and audio domains. Their modular, extensible architectures support both research-level experimental control and deployment in production-scale, user-interactive environments, outperforming traditional single-model approaches in both objective harmonization metrics and human evaluation (Jung et al., 2024, He et al., 8 May 2026).