We’ve spent the last few months developing HeartMuLa, an open-source music foundation model designed to bridge the gap between closed-source giants and the open community.
Why we built this: Existing models often struggle with lyric clarity in long compositions or require massive proprietary infrastructure. HeartMuLa focuses on high-fidelity, long-form music generation (up to 6 mins) with a specific emphasis on lyric intelligibility and multi-lingual support.
Technical Highlights:
1.HeartCodec: A new 12.5 Hz audio codec that achieves higher compression without losing musicality.
2.Architecture: Based on a hierarchical LLM approach, enabling fine-grained control over structure and style.
3.Efficiency: Optimized for consumer-grade GPUs, supporting local inference.
Everything is released under the Apache 2.0 license. We believe music AI should be hackable and transparent.
We’d love to hear your feedback on the audio quality and the architecture. Check out the demos and the code in the link above!
We’ve spent the last few months developing HeartMuLa, an open-source music foundation model designed to bridge the gap between closed-source giants and the open community.
Why we built this: Existing models often struggle with lyric clarity in long compositions or require massive proprietary infrastructure. HeartMuLa focuses on high-fidelity, long-form music generation (up to 6 mins) with a specific emphasis on lyric intelligibility and multi-lingual support.
Technical Highlights:
1.HeartCodec: A new 12.5 Hz audio codec that achieves higher compression without losing musicality.
2.Architecture: Based on a hierarchical LLM approach, enabling fine-grained control over structure and style.
3.Efficiency: Optimized for consumer-grade GPUs, supporting local inference.
Everything is released under the Apache 2.0 license. We believe music AI should be hackable and transparent.
We’d love to hear your feedback on the audio quality and the architecture. Check out the demos and the code in the link above!
Is fine-tuning / lora training supported? just toyed around with it, initial results are promising - more so than the recent Ace Step 1.5