WAIvePulse: Generate AI Music Locally — No Subscription, No Cloud, No Limits
Every AI music tool I looked at had the same problem: a monthly fee, a watermark, or a cap on how many songs you could generate. Suno, Udio, Udio — great tools, but the moment you start actually using them, you're watching a credit counter tick down.
So I built WAIvePulse. It runs on your machine, uses your GPU, and generates as many tracks as you have patience for — with no account, no API key, no subscription, and no data leaving your computer.
What It Does
You give it lyrics and a set of genre/style tags. It returns an MP3 with vocals. That's the core of it. But there's a lot more underneath:
- Web UI at
localhost:7860— dark-themed, no install beyond running the launcher - Job queue — submit multiple tracks and walk away while they process
- Real-time progress — live logs showing exactly where generation is in the two-phase pipeline
- Studio mode — stem separation powered by Demucs (Facebook Research) splits your track into 6 individual stems: vocals, drums, bass, guitar, piano, and other. Each has its own fader, mute, solo, EQ, reverb, and pan.
- Mix presets — Karaoke (vocals removed), Acappella (instruments removed), Drums Only
- BPM and key detection — automatic via librosa
- A-B looping — mark a section of the waveform and loop it for practice or editing
- 7 visualizer styles — Ring, Bars, Wave, Galaxy, Aurora, Particles, Scope. Fullscreen supported.
How It Works
The generation pipeline has two phases:
- Language model phase — HeartMuLa-oss-3B (a 3-billion parameter model) processes your lyrics and style tags and generates audio tokens autoregressively at roughly 1.5 tokens per second
- Codec decode phase — HeartCodec converts those audio tokens into a waveform at approximately 41 seconds per step
The FastAPI backend queues jobs in FIFO order and streams progress to the browser via Server-Sent Events, so you always know where you are in the process.
What You'll Need
This isn't a tool for a Chromebook. Requirements are real:
- GPU: NVIDIA with at least 10GB VRAM. CUDA is required — macOS is not supported for this reason.
- OS: Windows 10/11 or Linux
- Python: 3.10+
- Storage: ~21GB for model weights (the HeartMuLa model is ~15GB, HeartCodec is ~6.2GB)
Setup is automated. Run setup.bat on Windows or bash setup.sh on Linux. It handles Python, ffmpeg, CUDA verification, virtual environment creation, and model downloads. Then start.bat (or bash start.sh) to launch the server.
How Long Does Generation Take?
Honest answer: it takes a while. On a mid-range 12GB VRAM GPU:
- 30-second track: 10–12 minutes
- 1-minute track: 20–25 minutes
- 2-minute track: 40–50 minutes
- 5-minute track: 100–130 minutes
This is the tradeoff for running locally instead of on a cloud GPU farm. The job queue is there specifically so you can submit a batch, walk away, and come back to finished tracks.
Using the Lyric Markers
Structure your lyrics with section markers for better results:
[Intro]
Your intro lines here
[Verse]
Verse lyrics here
[Chorus]
Chorus lines here
[Outro]
Closing lines
For genre tags, the UI organizes them into 8 categories. Pick one per category for best results: Genre (required), Timbre, Gender (vocalist type or "no vocals" for instrumentals), Mood, Instrument, Scene, Region, and Topic.
The Studio
Once generation finishes, clicking Studio runs Demucs stem separation on the track. This takes a few more minutes but gives you a full DAW-style mixer with all 6 stems isolated. You can mute the vocals for a karaoke track, pull the drums forward, kill everything but the bass — whatever you need. Export the mix as a WAV when you're done.
Get It
WAIvePulse is free and open source on GitHub. If you have the hardware, there's no reason to keep paying monthly for AI music generation.
Comments