PhraseVAE and PhraseLDM: Latent Diffusion for Full-Song Multitrack Symbolic Music Generation

TL;DR

This work tackles full-song multitrack symbolic music generation by shifting from note-level tokens to phrase-level semantics. It introduces PhraseVAE, a 64-d latent compressor achieved via multi-query cross-attention and a progressive bottleneck, and PhraseLDM, a latent diffusion model that generates entire songs non-autoregressively in the phrase latent space. The framework enables efficient generation of up to 128 bars across multiple tracks with 45M parameters, leveraging length and structure conditioning to shape form while maintaining musical coherence and diversity. New evaluation metrics—PhraseFID, BarSSM, and memorization tests—demonstrate strong latent-space modeling, structure capture, and high novelty, suggesting phrase-level latent diffusion as a scalable, effective paradigm for long-form symbolic music generation.

Abstract

This technical report presents a new paradigm for full-song symbolic music generation. Existing symbolic models operate on note-attribute tokens and suffer from extremely long sequences, limited context length, and weak support for long-range structure. We address these issues by introducing PhraseVAE and PhraseLDM, the first latent diffusion framework designed for full-song multitrack symbolic music. PhraseVAE compresses an arbitrary variable-length polyphonic note sequence into a single compact 64-dimensional phrase-level latent representation with high reconstruction fidelity, allowing a well-structured latent space and efficient generative modeling. Built on this latent space, PhraseLDM generates an entire multi-track song in a single pass without any autoregressive components. The system eliminates bar-wise sequential modeling, supports up to 128 bars of music (8 minutes at 64 bpm), and produces complete songs with coherent local texture, idiomatic instrument patterns, and clear global structure. With only 45M parameters, our framework generates a full song within seconds while maintaining competitive musical quality and generation diversity. Together, these results show that phrase-level latent diffusion provides an effective and scalable solution to long-sequence modeling in symbolic music generation. We hope this work encourages future symbolic music research to move beyond note-attribute tokens and to consider phrase-level units as a more effective and musically meaningful modeling target.

Figures (9)

• Figure 1: Data representation and framework overview.
• Figure 2: PhraseVAE training stages.
• Figure 3: PhraseLDM model structure.
• Figure 4: Examples of bar-level Self-Similarity Matrices (SSMs) from a real song, a weakly structured generated song, and a structurally coherent generated song.
• Figure 5: Case studies of generated piano textures. The examples demonstrate idiomatic phrasing, structural coherence, stylistic diversity, and practical playability.