754,000+ skills indexed

vivekgoquestCOMMUNITY

Explore all skills by this author →

Verified:GitHub

heartmula

Generate full songs from lyrics and style tags using HeartMuLa open-source music models.

Updated Aug 22, 2026

Search Tags:#audio #gpu #music generation #lyrics #heartmula #open-source suno #ai songs

One-click install

npx skills add https://github.com/vivekgoquest/hermes-agent-stable --skill heartmula-vivekgoquest

Or copy as Structured Prompt for Agent▼

Please help me install this Agent Skill.
Skill: heartmula
Source: https://github.com/vivekgoquest/hermes-agent-stable/tree/main/optional-skills/creative/heartmula
Command: npx skills add https://github.com/vivekgoquest/hermes-agent-stable --skill heartmula-vivekgoquest

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Generating complete songs with vocals from text requires either paid services like Suno or complex local model setups. This Skill guides the installation, patching, and execution of HeartMuLa, an open-source music foundation model that produces full songs locally from lyrics and style tags. ## Core Features & Use Cases - Song Generation from Lyrics + Tags: Produce MP3 songs (48kHz stereo) conditioned on bracketed lyrics and comma-separated style tags using the HeartMuLa 3B/7B models and HeartCodec audio codec. - VRAM-Constrained Operation: Run on GPUs with as little as 8GB VRAM using lazy loading, or split models across multiple GPUs. - Dependency & Source Patching: Includes required fixes for transformers 5.x RoPE cache initialization and HeartCodec checkpoint loading. - Use Case: A musician wants an open-source Suno alternative to prototype song ideas locally. Provide lyrics with [Verse]/[Chorus] structure and tags like "piano,happy,wedding" to generate a 4-minute MP3 in roughly 4 minutes on a single GPU. ## Quick Start Use the heartmula skill to generate a song from my lyrics file and the tags "rock,energetic,guitar" and save it as an MP3.

Frequently Asked Questions about heartmula

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

FAQPage Schema

How do I generate a song from lyrics with AI locally?▼

Install the heartlib repository, download the HeartMuLa and HeartCodec checkpoints, then run run_music_generation.py with a lyrics file using [Verse]/[Chorus] tags and a comma-separated tags file. Output is a 48kHz stereo MP3.

What open-source alternative to Suno exists for song generation?▼

HeartMuLa is an Apache-2.0 family of music foundation models that generates full songs from lyrics and tags, comparable to Suno. It includes the HeartMuLa language model, HeartCodec audio codec, and lyrics transcription tools.

How much VRAM does HeartMuLa need for music generation?▼

HeartMuLa requires a minimum of 8GB VRAM with lazy loading enabled, which peaks around 6.2GB for the 3B model. 16GB or more is recommended, and multi-GPU setups can split the model and codec across devices.

Can HeartMuLa run on CPU or macOS?▼

CPU mode works with --mula_device cpu but is extremely slow, taking 30-60+ minutes per song with 12GB+ RAM. Triton GPU acceleration is unavailable on macOS, so Linux with CUDA is the supported GPU path.

Why does HeartMuLa fail with newer transformers versions?▼

Transformers 5.x breaks RoPE cache initialization because models load on meta device first, and HeartCodec checkpoints have mismatched scalar buffer shapes. Apply the documented patches to modeling_heartmula.py and music_generation.py.

Why are my style tags ignored during song generation?▼

Tag ignoring is a known upstream issue (#90) where lyrics tend to dominate the conditioning. Experiment with tag ordering and wording to influence the output style.

Related Skills

Explore More Design & Creative Skills→

music-lyrics

Extract timestamped lyrics from audio using faster-whisper.

composing-videos

Combine multiple videos or images into a single video with background audio.

sunoflow

Generate songs, separate stems, and convert audio via the Suno API.

songwriting-and-ai-music

Generates song lyrics, structure, and Suno AI music prompts with metatags.