HeartMuLa: A Family of Open Sourced Music Foundation Models

TL;DR

HeartMuLa presents an open-source family of music foundation models that unifies audio-text alignment, lyric recognition, music tokenization, and controllable long-form song generation. The core innovation is HeartCodec, a low-frame-rate ($f_l=12.5$ Hz) discrete-token tokenizer with $K=8$ codebooks of size $V=8192$, enabling efficient autoregressive modeling when paired with a hierarchical HeartMuLa generator built on a $\text{LLama}$-3.2 backbone. The framework introduces four-stage progressive training and Direct Preference Optimization (DPO) to achieve high musical quality, robust style control, and superior lyric intelligibility across multiple languages, with aggressive inference-speedups via KV-Cache, FlashAttention, and CUDA Graph. Extensive objective and subjective evaluations on HeartBeats benchmarks demonstrate competitive reconstruction and generation quality, while open-sourcing the weights and protocols supports reproducibility and community-driven advancement in multimodal music AI.

Abstract

We present a family of open-source Music Foundation Models designed to advance large-scale music understanding and generation across diverse tasks and modalities. Our framework consists of four major components: (1) HeartCLAP, an audio-text alignment model; (2) HeartTranscriptor, a robust lyric recognition model optimized for real-world music scenarios; and (3) HeartCodec, a low-frame-rate (12.5 Hz) yet high-fidelity music codec tokenizer that captures long-range musical structure while preserving fine-grained acoustic details and enabling efficient autoregressive modeling; (4) HeartMuLa, an LLM-based song generation model capable of synthesizing high-fidelity music under rich, user-controllable conditions (e.g., textual style descriptions, lyrics, and reference audio). In addition, it provides two specialized modes: (i) fine-grained musical attribute control, which allows users to specify the style of different song sections (e.g., intro, verse, chorus) using natural language prompts; and (ii) short, engaging music generation, which is suitable as background music for short videos. Lastly, HeartMuLa improves significantly when scaled to 7B parameters. For the first time, we show that a Suno-level, commercial-grade system can be reproduced using academic-scale data and GPU resources. We expect these foundation models to serve as strong baselines for future research and to facilitate practical applications in multimodal content production.

Figures (4)

• Figure 1: Overall comparison of HeartMuLa-oss-3B with existing music foundation models.
• Figure 2: An illustration of our proposed HeartCodec. Left, middle, right are semantic-rich encoder, ultra-low frame rate compressor and high-fidelity reconstruction decoder, respectively.
• Figure 3: HeartMuLa Architecture
• Figure 4: Four-Stage Progressive Training Paradigm