Pushing the Frontier of Full-Song Generation: Hierarchical Autoregressive Planning Meets Flow-Matching Rendering

· Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Generative Audio & Music · Depth: Expert, quick

Summary

A new unified song generation framework has been introduced, capable of producing high-quality, full-length music from various inputs including lyrics, text descriptions, and musical attributes. This framework supports three distinct tasks: Lyrics-to-Song Generation, Instrumental Music Generation, and Cover Song Generation, which reinterprets existing songs with different styles while preserving melodic content. Architecturally, the system integrates a semantic-aware tokenizer that encodes audio into 8-codebook RVQ tokens, a hybird-LM for hierarchical autoregressive audio-token modeling, and FullDiT for full-song flow matching in a continuous VAE latent space. A two-level melody module specifically guides cover song generation by extracting and discretizing melody cues. The framework also incorporates DPO, GRPO, and OPD as reward-based post-training strategies for hybird-LM, and flow-based GRPO for FullDiT, to enhance musicality and rendering quality. Experimental results demonstrate competitive performance on a multilingual automatic benchmark and the Artificial Analysis Music with Vocals leaderboard.

Key takeaway

For Machine Learning Engineers developing advanced music generation systems, this framework presents a robust, multi-task architecture for full-song synthesis. You should consider integrating hierarchical autoregressive planning with flow-matching techniques, like FullDiT, to achieve high-fidelity audio. Furthermore, applying reward-based post-training strategies such as DPO or GRPO can significantly enhance the musicality and rendering quality of your generated outputs across diverse tasks like lyrics-to-song or cover generation.

Key insights

A unified framework combines hierarchical autoregressive planning and flow-matching for high-quality, full-song generation from diverse inputs.

Principles

Method

Audio is tokenized into 8-codebook RVQ tokens. hybird-LM performs hierarchical autoregressive modeling. FullDiT applies flow matching in VAE latent space. A melody module guides cover songs, and DPO/GRPO/OPD refine models.

In practice

Topics

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.