PCo-Mamba: Phase-Coherent Complex-Domain State Space Models for Generative Music Synthesis
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
High-fidelity music generation models based on Transformer and diffusion architectures suffer from two intrinsic bottlenecks: (1) the quadratic O(n 2 ) time complexity of selfattention, which entails prohibitive computational costs for high-resolution, long-duration audio sequences; and (2) the reliance on conventional amplitude-spectrum modeling, which neglects critical phase information and leads to severe phase distortion in reconstructed signals.To address these challenges, we propose PCo-Mamba, a novel phase-coherent complex-domain state-space framework.The proposed system integrates four synergistic innovations: 1. Complex-Domain State Space Models (SSMs), which jointly model the real and imaginary components in complex-spectral space to ensure fundamental phase coherence through complex-valued operators; 2. Band-Adaptive Mamba Encoding (Band-Adaptive CBE), which utilizes a selective scan mechanism with frequency-dependent step sizes (Δ) to achieve precise, multi-scale acoustic dynamics; 3. Retrieval-Augmented Guidance (RAG), which injects non-parametric high-fidelity timbre priors to mitigate spectral hallucinations in high-frequency regimes; 4. Flow Matching Reconstruction, which leverages ODE-based probability path optimization to achieve smoother distribution transitions and accelerated inference compared to traditional diffusion processes.Experimental results on the MAESTRO v3.0.0 piano dataset demonstrate that PCo-Mamba achieves state-of-the-art (SOTA) performance across multiple metrics.Compared to leading baselines such as MusicGen and AudioLDM 2, our system achieves a 3.7 dB improvement in Signal-to-Distortion Ratio (SDR) and a 36% reduction in Fréchet Audio Distance (FAD).Perceptual evaluations via MUSHRA tests further corroborate that PCo-Mamba significantly outperforms existing architectures in terms of keyboard touch sensitivity, phase authenticity, and long-term coherence.Furthermore, the linear computational complexity of PCo-Mamba enables the generation of ultra-long sequences, paving the way for the next generation of high-performance interactive music synthesis.
Fields of Study & Subjects
Music and Audio Processing
Music Technology and Sound Studies
Speech Recognition and Synthesis
Publication details
- Journal / source
- Traitement du signal
- Volume
- 43
- Issue
- 1
- Publication date
- 28 February 2026
- Language
- EN