Embedding Alignment in Code Generation for Audio¶

Conference: NeurIPS 2025 arXiv: 2508.05473
Code: None
Area: Code Intelligence Keywords: code generation, audio embedding, contrastive learning, cross-modal alignment, live-coding

TL;DR¶

A dual-MLP + InfoNCE contrastive learning framework is proposed to align code embeddings (distilroberta-base) and audio embeddings (wav2vec2) into a shared space, enabling LLM-based code generation pipelines to infer musical similarity directly from code without compilation or execution. CKA improves from 0.090 to 0.590.

Background & Motivation¶

Live-coding scenario: Performers must write music-generating code (e.g., Sonic Pi) in real time under time pressure and in front of an audience. LLM-assisted code generation can reduce syntactic burden, allowing creators to focus on high-level musical ideas.
Core limitation: Existing LLM code generation models evaluate candidate code using text similarity metrics (e.g., BLEU, edit distance), but textual similarity does not imply audio similarity — two textually similar programs may produce drastically different sounds, and vice versa.
Key observation: Computing code-to-code and audio-to-audio embedding distances on 27 Sonic Pi tutorial examples yields a Pearson correlation of only 0.0159 (\(p=0.677\)) and a Spearman correlation of only 0.0409 (\(p=0.445\)), indicating virtually no linear or rank-order relationship between the raw embedding spaces.
Parameter perturbation experiment: Minor code modifications (sleep, amplitude, bpm) keep code embedding similarity above 0.990, while audio embedding similarity varies more widely (below 0.975), with no consistent pattern across parameter types — demonstrating that the alignment mapping is non-trivial.

Method¶

Data Construction¶

Using 27 Sonic Pi tutorial code examples, parameters (synths, samples, notes, attack/release, amp, sleep, effects, etc.) are randomized via the Jinja templating engine.
500 distinct Sonic Pi code files are generated, yielding 13,500 code–audio paired samples in total.
Code embeddings are produced by distilroberta-base; audio embeddings by Meta's wav2vec2. Audio is rendered at 120 BPM over 9 bars.

Symmetric Dual-MLP Architecture¶

Two independent MLPs process code embeddings \(\texttt{MLP}_c\) and audio embeddings \(\texttt{MLP}_a\), respectively.
Each MLP consists of \(L\) linear layers with hidden dimension \(d_{\text{hidden}}\), BatchNorm, and GELU activations.
Pre-trained embeddings are projected into a shared space: \(c_i = \texttt{MLP}_c(c_i^0)\), \(a_i = \texttt{MLP}_a(a_i^0)\).
A symmetric MLP design is chosen over attention mechanisms, as the goal is efficient mapping without introducing modality bias.

InfoNCE Contrastive Learning¶

Given \(N\) aligned code–audio embedding pairs \(\{(c_i, a_i)\}_{i=1}^N\) in a batch, cosine similarity is computed as:

\[\text{sim}(c_i, a_j) = \frac{c_i^\top a_j}{\|c_i\| \cdot \|a_j\|}\]

The InfoNCE loss pulls positive pairs closer and pushes negative pairs apart:

\[\mathcal{L}_i = -\log \frac{\exp(\text{sim}(c_i, a_i) / \tau)}{\sum_{j=1}^N \exp(\text{sim}(c_i, a_j) / \tau)}\]

\(\tau\) is a temperature hyperparameter controlling the sharpness of the similarity distribution.
This self-supervised approach requires no explicit annotations and learns alignment solely from code–audio pairing.

Evaluation Metrics¶

Metric	Type	Description
CKA (Centered Kernel Alignment)	Structural similarity	Invariant to orthogonal transformations and isotropic scaling; captures nonlinear structural similarity
CCA (Canonical Correlation Analysis)	Linear correlation	Measures maximum linear correlation between two sets of multivariate variables
Jaccard / overlap@k	Neighborhood consistency	Whether nearest neighbors in code space correspond to nearest neighbors in audio space
Spearman / Pearson	Rank / linear correlation	Degree of correlation in distance rankings

Key Experimental Results¶

Hyperparameter Tuning (24 configurations, averaged over 5 runs)¶

Setting	CKA	CCA
Pre-alignment baseline	0.090	0.140
Best config (CKA)	0.590	—
Best config (CCA)	—	0.902

Both metrics achieve a 6× or greater improvement.

Three Code Completion Scenarios¶

Scenario	Method	Jaccard	overlap@3	Spearman	Pearson
melody	Raw baseline	0.20	0.33	0.21	0.18
melody	Ours	0.34	0.47	0.16	0.07
drum	Raw baseline	0.00	0.00	—	-0.25
drum	Ours	0.16	0.27	-0.05	-0.12
bass	Raw baseline	0.20	0.33	0.24	0.21
bass	Ours	0.50	0.67	0.44	0.46

Key Findings¶

Consistent neighborhood metric improvements: Jaccard and overlap@3 surpass the raw baseline in all three scenarios.
Largest gain on drum: The baseline completely fails (Jaccard=0); alignment raises it to 0.16/0.27.
Strongest performance on bass: Jaccard 0.50, overlap@3 0.67, with significant improvements in rank correlation.
No audio compilation required: All inference is performed directly on code embeddings, avoiding the computational overhead of audio rendering.
UMAP visualization: Before alignment, code (blue) and audio (orange) embeddings are completely separated; after alignment, the two modalities overlap in the embedding space, with semantically related code–audio pairs clustering in nearby regions.

Highlights & Insights¶

Novel problem formulation: This work is the first to systematically study cross-modal alignment between code and audio embeddings, filling a gap in the niche but active field of creative coding / live-coding.
Lightweight yet effective: Significant alignment is achieved with only dual MLPs and InfoNCE, without complex cross-modal Transformer architectures.
Practical value: Post-alignment, audio similarity can be inferred directly from code embeddings, endowing code completion tools with "music-aware" capabilities and enabling LLM-assisted live-coding systems to generate more diverse and perceptually meaningful candidate code.
Transferable methodology: The approach of first conducting negative experiments (demonstrating the absence of correlation in raw spaces) to motivate the alignment model provides a clear and replicable argumentative structure.

Limitations & Future Work¶

Limited data scale: The dataset is expanded from only 27 Sonic Pi tutorial templates, resulting in insufficient diversity in musical style and code structure.
Sonic Pi only: Generalization to other music programming languages (SuperCollider, TidalCycles, Strudel) is not validated.
Unstable rank correlation metrics: Spearman/Pearson do not improve consistently in the melody and drum scenarios, indicating that alignment remains insufficient for fine-grained ranking.
Not integrated into a real code completion system: Validation is conducted only in offline experiments and has not yet been embedded into an actual LLM code assistant pipeline.
Embedding model selection: distilroberta-base is a general-purpose text model not optimized for code semantics; replacing it with CodeBERT or similar models may yield better results.
Audio representation: wav2vec2 is primarily designed for speech; its music representation may be imprecise. Music-specific embeddings such as CLAP could be considered.

Unlike music–text joint embedding works such as MuLan (Huang et al., 2022), this paper focuses on the unique cross-modal pair of code–audio.
The cross-modal alignment approach (lightweight MLP + contrastive learning) is transferable to other scenarios where "code produces non-textual outputs," such as visualization code → images or simulation code → motion trajectories.
This work offers insights for subsequent research in creative AI / AI-assisted music: how to equip LLMs with awareness of output modalities during code generation.
The methodological contributions are transferable to related problems.

Rating¶

Novelty: ⭐⭐⭐⭐ Novel perspective on code–audio alignment
Experimental Thoroughness: ⭐⭐⭐⭐
Writing Quality: ⭐⭐⭐⭐
Value: ⭐⭐⭐⭐ Specific value for the audio generation domain