music-gen

name	music-gen
description	Generate original songs with lyrics and vocals. Use when users ask to sing, create a song, make music, compose, or generate audio. Supports style prompts, custom lyrics, duration control, and instrumental mode.
allowed-tools	["Bash"]

Music Generation Skill

Generate original songs using ACE-Step 1.5 on remote 3090 GPU. Produces full songs with vocals, instruments, and lyrics in ~40 seconds.

When to Use

• User says "sing me a song", "make a song", "create music"
• User provides lyrics and asks for a song
• User describes a music style they want to hear
• User says "generate a song about..."

Command

alma sing generate "style description" [--lyrics "lyrics text"] [--duration 60] [--instrumental]

Parameters

Flag	Default	Description
"style description"	required	Music style/genre prompt (e.g. "dreamy lo-fi chill hop, female vocal")
--lyrics "..."	empty	Song lyrics with structure tags like [verse], [chorus], [bridge]
--duration N	60	Duration in seconds (10-600)
--instrumental	false	Generate without vocals

Output

Prints the saved MP3 file path to stdout. Send it with:

alma send file <chatId> <path>

Examples

Simple song

alma sing generate "upbeat pop rock, male vocal, energetic drums"

Song with custom lyrics

alma sing generate "dreamy lo-fi chill hop, female vocal, warm piano" --lyrics "[verse]\nMidnight glow on my screen\nCoffee steam like a dream\n\n[chorus]\nLet the music carry me\nThrough the night so peacefully"

Instrumental track

alma sing generate "epic orchestral cinematic, strings and brass" --duration 120 --instrumental

Short jingle

alma sing generate "cute happy ukulele jingle" --duration 15

Style Prompt Tips

• Include genre (pop, rock, jazz, lo-fi, classical, EDM, hip hop...)
• Include vocal type (female vocal, male vocal, choir...)
• Include instruments (piano, guitar, drums, synth, strings, brass...)
• Include mood (dreamy, energetic, melancholic, upbeat, dark, romantic...)
• Example: "melancholic acoustic folk, female vocal, fingerpicked guitar, autumn evening vibe"

Lyrics Format

Use structure tags for song sections:

[verse]
First verse lyrics here
Line two of verse

[chorus]
Chorus lyrics here
Repeated refrain

[bridge]
Bridge section
Something different

[outro]
Ending lyrics

Supported Languages

50+ languages for lyrics including English, Chinese, Japanese, Korean, Spanish, French, etc.

Technical Notes

• Backend: ACE-Step 1.5 on NVIDIA RTX 3090 (24GB VRAM)
• LM: 0.6B parameter model with PyTorch backend
• Generation time: ~35-40s for 60s audio
• Audio format: MP3
• The API server auto-starts if not running (first call takes ~90s extra for model loading)
• ACE-Step and ComfyUI share the same GPU — ComfyUI will be stopped when music generation starts
• Output saved to ~/.config/alma/music/

Troubleshooting

If generation fails:

1. Check if the 3090 is reachable: ssh yetone@10.0.0.207 'echo ok'
2. Check ACE-Step logs: ssh yetone@10.0.0.207 'tail -30 /tmp/acestep.log'
3. Check VRAM: ssh yetone@10.0.0.207 '/usr/lib/wsl/lib/nvidia-smi'
4. If VRAM is full, kill other processes: ssh yetone@10.0.0.207 'pkill -f comfyui; pkill -f acestep'