Type a vibe. Get a whole song. On your own Mac. No cloud, no subscription, no API keys.
You write something like "sunset reggae with steel pan and a male vocalist who sounds like Burning Spear, 78 bpm" — Song Forge writes lyrics for you (or you can write your own), an AI composes the music and the vocals, and a couple minutes later you've got a finished track in your Library. Drop your own voice sample in and it'll re-sing the song in your voice — or your kid's, or anyone you've recorded.
Everything happens on your Mac. Nothing ever leaves it.
A full walkthrough on YouTube — describing a vibe, generating the song, swapping in a cloned voice. Click the thumbnail above to watch.
- You describe a song — pick a genre preset, mix in mood / era / tempo / reference artists, fine-tune the vocal arrangement.
- The AI writes the lyrics for you in the right style (or paste your own —
[verse]/[chorus]tags work). - It generates the song — music and vocals — in about 15 seconds on a beefy Mac (a couple minutes on a normal one).
- Library keeps every track you've made. Play, rename, save the WAV, delete.
- Voice swap any track into a cloned voice. Drop a 10-second sample of someone speaking into the voices folder, and Song Forge will re-sing the whole song in their voice. There's even a "group of kids" effect that layers 4 pitched copies for a children's choir feel.
Apple Silicon Mac. Sorry Intel folks — the music model uses Apple's MLX framework, which needs an M-series chip.
| Your Mac | What you'll get |
|---|---|
| M1/M2 with 16 GB RAM | Lyrics + music will work, but you'll have to skip Gemma (the lyric writer) or run a smaller model. Songs take ~2 min. Voice swap is tight on memory. |
| M2/M3 Pro/Max, 32 GB RAM | Sweet spot for most people. Run everything comfortably. Songs in 30–60 seconds. |
| M3/M4 Max, 64 GB+ RAM | Snappy. Songs in 20–30 seconds. Multiple jobs at once if you want. |
| M5 Max, 128 GB RAM (my machine) | About 15 seconds for a 2-minute song. Voice swap in under a minute. Everything stays buttery. |
Disk: Plan for ~30 GB free for the AI models the first time you set it up. Generated songs are ~20 MB each — get a hundred and you've used 2 GB.
Microphone (optional): Only needed if you want to record your own voice samples for the voice-swap feature.
Internet: Required once to download the models. After that — totally offline. Take it on a plane.
1. Install LM Studio (for the lyric writer) → lmstudio.ai, pull a Gemma model, hit "Start Server."
2. Set up the music engine — clone ACE-Step somewhere on your machine. The supervisor script expects it.
3. Install seed-vc and demucs for the voice-swap pipeline (pip install seed-vc demucs).
4. Run it:
bash forge_supervisor.shOpen http://localhost:8767 and hit FORGE A SONG.
| Component | Role |
|---|---|
| ACE-Step 1.5 | The music brain. Generates instrumentation + vocals from your style description. (MLX backend on Apple Silicon.) |
| Gemma 4 31B (via LM Studio) | The lyricist. Writes verse/chorus lyrics that fit your genre. |
| Demucs | Splits a finished song into vocals + instrumental. |
| seed-vc (Plachta) | Zero-shot voice cloning. Give it a 10-second sample of someone, it'll re-sing any vocal in their voice. |
| ffmpeg | Mixes the new vocals back over the instrumental, loudness-normalized to −14 LUFS so it sounds pro. |
forge_server.py |
The Python web server that ties it all together. Runs on port 8767. |
index.html |
The whole web UI — liquid-glass theme, drifting cosmic colors, big chunky buttons. |
- Generate the same lyrics in five different genres — paste the same
[verse]/[chorus]block, change the style preset, hit MAKE. You'll have five wildly different versions of your song in two minutes. - Drop a sample of yourself reading a paragraph into the voices folder. Voice-swap any song to your voice. Yes, it's surreal.
- Use the "group of kids" effect for an instant children's-choir version of any chorus.
- Type a banned-phrases list (concrete jungle, one love, irie, etc.) — Gemma will refuse to use them in lyrics.
- Open Fine Tune — control era, mood, vocal character, tempo, reference artists, extra instruments. Each one stacks cleanly into the style description.
This is a personal project. No license assigned, no support promised, no warranty. Built it for myself and put it up because someone might learn from it.
Models are large. Setup takes a minute. But once it's running it's just… yours. Forever. Offline. No one tracking what you make. No subscription that disappears when the company pivots.
Have fun.