heartmula

Generate full song audio from lyrics and tags using HeartMuLa and HeartCodec.

Updated Apr 10, 2026

Search Tags:#ai #audio #huggingface #music-generation #heartmula #heartcodec #lyrics-to-song

Ready to run

npx skills add https://github.com/KarlinskyS/hermesSkills --skill heartmula-karlinskys

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: heartmula
Source: https://github.com/KarlinskyS/hermesSkills/tree/main/media/heartmula
Command: npx skills add https://github.com/KarlinskyS/hermesSkills --skill heartmula-karlinskys

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

HeartMuLa enables creators to turn lyrics and tag-based prompts into complete, high-fidelity songs using open-source models, removing the need for manual music production or proprietary services.

Core Features & Use Cases

Lyrics-conditioned generation: Produce full songs from structured lyrics and tag lists (genre, mood, instruments).
High-quality codec: Uses HeartCodec for fidelity-preserving audio reconstruction and HeartTranscriptor for optional lyrics transcription.
Flexible deployment: Local GPU or CPU execution with lazy-loading, multi-GPU splitting, and model checkpoint management.
Use Case: A songwriter generates demo tracks in multiple languages from a single lyrics file and iterates rapidly by adjusting tags and sampling parameters.

Quick Start

Generate a 90-second song from the provided lyrics and tags and save the result as a 48kHz MP3.

Frequently Asked Questions about heartmula

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I generate full songs from lyrics and genre tags?▼

To generate full songs from lyrics and tags, you provide structured lyrics and tag lists to condition the audio synthesis. The system processes these inputs using the HeartMuLa model to produce complete, high-fidelity song audio tracks.

Can I run lyrics-to-song generation locally without a GPU?▼

Yes, you can run lyrics-to-song generation locally without a GPU by using the CPU fallback. The system supports flexible deployment with PyTorch on CPU, though utilizing CUDA with a GPU provides significantly faster audio synthesis.

What is lyrics-conditioned audio synthesis and how does it work?▼

Lyrics-conditioned audio synthesis is the process of generating music directly from text lyrics and descriptive tags. It works by using the HeartCodec to preserve audio fidelity while translating lyrical structure and mood tags into high-quality song waveforms.

Does this music generation workflow support multilingual lyrics?▼

Yes, this music generation workflow supports multilingual lyrics for audio synthesis. You can generate demo tracks from a single lyrics file containing multiple languages by adjusting the sampling parameters and tags.

Do I need Python 3.10 and PyTorch for audio synthesis?▼

Yes, you need Python 3.10 and PyTorch with CUDA for audio synthesis. These are required dependencies for managing model checkpoints and executing the HeartMuLa and HeartCodec models during song generation.

What are the limitations of generating 48kHz MP3 songs from text?▼

Limitations of generating 48kHz MP3 songs from text include hardware constraints for local execution and the need for manual tag adjustments to refine output. Users must manage model checkpoints and rely on lazy-loading to optimize memory during synthesis.