HeartMuLa

Four models make up the HeartMuLa family, an open music foundation stack: the HeartMuLa language model generates full songs conditioned on multilingual lyrics and style tags, HeartCodec compresses audio into 12.5 Hz tokens so long tracks stay tractable for the language model, HeartTranscriptor performs Whisper-based lyrics transcription, and HeartCLAP aligns music and text for retrieval and scoring. The released 3B generation model runs at roughly real-time speed on a single GPU, with lazy loading to fit smaller cards and multi-GPU placement for headroom, and a 7B variant is planned. Install is git clone plus pip on Python 3.10, with checkpoints pulled from Hugging Face or ModelScope; a January 2026 license update moved the whole project and its model weights to Apache 2.0, clearing commercial use. As one of the few serious open rivals to ACE-Step for lyric-conditioned song generation, with the codec and transcription pieces usable standalone, it has become a common base for open music research and tooling, sitting near 3.7k GitHub stars.