Noises 🎵

A local AI music generator that lives on your desktop. No subscriptions, no cloud wait times, just noise.

Noises is a desktop app built to make generating AI music, samples, and loops effortless. It uses Python to run AI models entirely on your own hardware.

It currently supports two generation modes:

• Loops — Powered by Stable Audio Open (Stability AI). Best for short loops, textures, and samples. Supports negative prompts, BPM, and key selection.
• Full Songs — Powered by ACE-Step. Generates complete songs with vocals, lyrics, and structure tags. Supports up to 4 minutes of audio.

Features

• 🎹 Loops & Samples: Create short loops at any BPM and Key using Stable Audio Open.
• 🎤 Full Songs with Vocals: Generate complete songs with lyrics, genre tags, and structure using ACE-Step.
• 🏠 100% Local: Everything runs on your own GPU. Your prompts and music stay private.
• 🎛️ Full Control: Tweak inference steps, guidance scale, scheduler type, CFG mode, seed, and more.
• ✨ Prompt Enhancement: One-click modifiers for clean/hi-fi, lo-fi, cinematic, aggressive, and other styles.
• 💾 Presets: Save and recall your favorite prompt configurations.
• 🎯 Drag & Drop: Drag generated audio directly into your DAW.
• 📂 Organized Output: All generated files are saved to ~/Music/Noises/ (loops and songs in separate folders).

How to Run it (The Easy Way)

1. Go to the Releases page.
2. Download the installer (Noises_Setup.exe).
3. Run it, install it, make noise.

Requirements: An NVIDIA GPU (RTX 3060 8GB or better recommended). No manual CUDA installation needed — Noises handles everything automatically.

First-run setup: On first launch, the installer will download and cache PyTorch with CUDA support (~2.5 GB) to C:\ProgramData\Noises\torch_runtime\. This is a one-time step. AI models (~3 GB for Stable Audio, ~3.5 GB for ACE-Step) are also downloaded on first use and cached in ~/.cache/huggingface/. Everything runs offline after that.

Troubleshooting: If setup fails, check %LOCALAPPDATA%\Noises\setup.log for details. If something breaks, please open an issue!

Task Manager & Processes

When running Noises, you will see a few different processes in your Task Manager. This is normal!

• Noises.exe: The main application window and lightweight controller.
• backend.exe: The heavy lifter. This is the Python engine running the AI models. It runs as a separate process to keep the UI smooth.
• WebView2 / msedgewebview2.exe: These processes handle the rendering of the user interface (similar to Chrome/Edge tabs).

Tech Stack

Layer	Technologies
Frontend	React, Tailwind CSS, Framer Motion, Lucide Icons
Backend	Python (FastAPI, Uvicorn), PyTorch, HuggingFace Diffusers, ACE-Step
Desktop	Rust (Tauri v2), Sidecar process management
Audio	soundfile, torchaudio, numpy

Building from Source

If you want to hack on it yourself:

Prerequisites

• Node.js 18+
• Python 3.12 (MS Store or python.org — must be Python 3.12 specifically, as the packaged backend is built against it)
• Rust (installed via rustup)
• System Requirements:
  • GPU: NVIDIA GeForce RTX 3060 (8GB VRAM) or better recommended.
  • Minimum: NVIDIA GPU with 6GB VRAM (Stable Audio). ACE-Step uses CPU offloading to fit within ~8GB peak VRAM.
  • Disk Space: ~11 GB free — ~2.5 GB for PyTorch (%ProgramData%\Noises\torch_runtime) + ~8 GB for models (~/.cache/huggingface).

Setup

# 1. Clone the repo
git clone https://github.com/yourusername/noises.git
cd noises

# 2. Create a Python virtual environment
python -m venv .venv
# Windows:
.venv\Scripts\activate
# Linux/macOS:
# source .venv/bin/activate

# 3. Install Python dependencies
pip install -r backend/requirements.txt

# 4. Install frontend dependencies
cd frontend
npm install
cd ..

# 5. Run in dev mode (from root folder)
npm run tauri dev

Output Location

Generated audio is saved to:

~/Music/Noises/
├── samples/
│   ├── loops/       # Stable Audio output (loop_120bpm_C_minor_001.wav, ...)
│   └── oneshots/    # ACE-Step output (song_001.wav, ...)

How It Works

1. Tauri starts the native window and spawns the Python backend as a sidecar process.
2. The frontend sends generation requests to the backend over HTTP (127.0.0.1:8000).
3. The backend loads the requested model on demand and unloads it after generation to free VRAM.
4. Generated audio is post-processed (normalization, fades) and saved as WAV files.
5. The frontend reads the file via Tauri's FS plugin and plays it in the browser audio element.

Credits

Big thanks to the open source community for making this possible:

• ACE-Step for full song generation with vocals
• Stability AI for Stable Audio Open
• Tauri for the desktop framework
• HuggingFace for model hosting and Diffusers

License

MIT License. Do whatever you want with the code, just credit me!

The generated audio is subject to the licenses of the respective models. Check the ACE-Step and Stable Audio Open repos for commercial use details.