Language: English · Русский · 中文 · 日本語 · Español · Français · Deutsch · Português · 한국어

---

Stable Audio Insight

By Oscar Lumiere

Quick start

1. Register at https://huggingface.co (skip if you already have an account)
2. Open https://huggingface.co/stabilityai/stable-audio-open-1.0 → click Agree and access repository
3. At https://huggingface.co/settings/tokens → New token → Type: Read → Create
4. Run install.bat → pick UI language → paste token → pick translator. Done.

Then run.bat opens http://127.0.0.1:7860.

Self-contained Windows app for Stable Audio Open 1.0 by Stability AI. Built for sound design, game audio, video editing and music sketches.

What's inside

• 9-language UI (English, Russian, Chinese, Japanese, Spanish, French, German, Portuguese, Korean) — switcher in the Gradio footer changes everything on the fly
• Multilingual prompts — write in Cyrillic / CJK / Arabic / Hebrew / Greek / Thai / Devanagari, gets auto-translated to English before generation. Two translator options at install: opus-mt-mul-en (light, ~300 MB) or nllb-200-distilled-600M (heavy, ~2.4 GB)
• ~217 ready presets in 15 categories — Footsteps, Impacts, Movement, UI, Weapons, Ambience, Vehicles, Nature, Music, Cinematic, Magic, Sci-Fi, Horror, Animals, Crowd & Voice
• Multi-variation generation — 1 to 4 audio outputs per click (num_waveforms_per_prompt), scheduler picker (Default / DPM++ 2M / Euler), random seed button, re-roll button
• Session history — last 10 generations as collapsible mini-players
• Batch mode — paste many prompts (one per line), each saved straight to outputs/Saved/
• ZIP export of outputs/Saved/ in one click
• Game-ready output — slugified filenames (wood_floor_footsteps_01.wav) with auto-incremented per-folder counters; WAV INFO chunk metadata carries title (full prompt), seed, duration, steps, cfg, sample rate, negative — read by Reaper, Audition, Ableton, ffprobe
• GPU/CPU live toggle plus RunOnCPU.bat for low-VRAM systems
• Per-session log file in logs/app-YYYYMMDD-HHMMSS.log with full tracebacks on errors
• One-file installer (install.bat) — silently sets up Python 3.11.9 inside the project folder, picks torch+cu128 or CPU torch via nvidia-smi auto-detect, downloads model and translator. Truly portable: deleting the project leaves zero traces in the system
• Update path — update.bat refreshes deps and re-syncs models

Requirements

• Windows 10 / 11 (tested on Windows 11 Pro)
• Python 3.11.x — https://www.python.org/downloads/release/python-3119/ (during install, check Add Python to PATH)
• NVIDIA GPU with CUDA 12.8 (tested on RTX 5080, 16 GB VRAM). For other GPUs, change the PyTorch index in install.bat / setup.bat (e.g. cu121 for CUDA 12.1).
• 15–20 GB free disk during install (after install: ~10–13 GB resident):
  • ~5 GB — Stable Audio Open weights
  • ~4–5 GB — Python + PyTorch CUDA + dependencies
  • 0.3–2.4 GB — translator (your choice)
  • ~5 GB — HuggingFace + pip caches + install temp files (mostly recoverable later)
• HuggingFace account + read token (only required for the initial download)

Do I need a HuggingFace token?

Once — only to download Stable Audio weights. After that the app runs offline from the local hf-cache/.

Stable Audio Open 1.0 is a gated model — Stability AI requires a registered account and acceptance of the model license before download. So:

1. Register on https://huggingface.co (if you don't already have an account)
2. Open the model page and accept the license: https://huggingface.co/stabilityai/stable-audio-open-1.0 (click Agree and access repository — without this the download returns 403)
3. Create a read token: https://huggingface.co/settings/tokens
4. install.bat will ask you to paste the token at the right moment

The translator models (opus-mt-mul-en and nllb-200-distilled-600M) are not gated and need no token.

Install (recommended path)

A single file does everything in one go:

install.bat

What it does:

1. Picks the prompts language (1–9: English, Russian, Chinese, Japanese, Spanish, French, German, Portuguese, Korean)
2. Verifies that Python 3.11 is on PATH (py -3.11)
3. Creates .venv and upgrades pip / wheel
4. Installs PyTorch (CUDA 12.8) and the rest of requirements.txt
5. Asks for the HF token (or skips if you're already logged in)
6. Asks which translator to install:
   • 1. LIGHT — Helsinki-NLP/opus-mt-mul-en
     • ~300 MB, fast, Apache 2.0 (commercial use OK)
     • Medium quality, sometimes mistranslates music terms
   • 2. HEAVY — facebook/nllb-200-distilled-600M
     • ~2.4 GB, slower, CC-BY-NC 4.0 (non-commercial only)
     • Higher quality, recommended for non-English
7. Saves the choice to hf-cache/translator.cfg
8. Downloads Stable Audio (~5 GB) and the chosen translator into hf-cache/

Run

run.bat

A browser tab opens at http://127.0.0.1:7860.

Switch the translator after install

Quickest path:

1. Open hf-cache/translator.cfg and change to light or heavy
2. Run download.bat — it will fetch the chosen one if it's not in cache
3. run.bat

Or delete hf-cache/translator.cfg and run install.bat again — it will ask anew and will not re-download dependencies or Stable Audio (already in cache).

Step-by-step install (if something goes wrong)

install.bat is a wrapper around four smaller scripts; each can be run alone:

Script	What it does
setup.bat	Creates .venv and installs dependencies
login.bat	HuggingFace authentication (hf auth login)
download.bat	Downloads Stable Audio + the translator from hf-cache/translator.cfg
run.bat	Launches the Gradio web UI

Project layout

.
├── app.py              # Gradio UI + generation
├── translator.py       # Translator module (light / heavy)
├── download_model.py   # Model downloads from HuggingFace
├── requirements.txt    # Python dependencies
├── install.bat         # Single-file installer (multilingual)
├── setup.bat           # Just venv + dependencies
├── login.bat           # Just HF auth
├── download.bat        # Just downloads
├── run.bat             # Launch the web UI
├── hf-cache/translator.cfg  # light / heavy (created by install.bat)
├── README.md           # main docs (English)
├── docs/i18n/          # localized README in 8 other languages
├── LICENSE             # MIT — for the source code in this repo
├── NOTICE.md           # Licenses of the models used
├── hf-cache/           # Local HF cache (NOT in git)
└── outputs/            # Saved audio (NOT in git)
    ├── intermediate/   # When the "Save all" checkbox is on
    └── Saved/          # When you press the "Save" button

Environment variables

You can override these before running run.bat:

Variable	Default	Purpose
HOST	127.0.0.1	Gradio host
PORT	7860	Gradio port
SHARE	0	1 = create a public gradio.live URL
STABLE_AUDIO_MODEL	stabilityai/stable-audio-open-1.0	Override the model

Licenses

This source code is MIT (see LICENSE).

The models each have their own licenses (see NOTICE.md for full text):

Model	License	Commercial use
Stable Audio Open 1.0	Stability AI Community	Up to $1M annual revenue: yes; above: Enterprise license required
opus-mt-mul-en (light)	Apache 2.0	Yes
NLLB-200-distilled-600M (heavy)	CC-BY-NC 4.0	No, non-commercial only

If you intend to use this commercially, pick the light translator at install.

Tested on

• Windows 11 Pro
• Python 3.11.9
• NVIDIA RTX 5080 (16 GB VRAM)
• CUDA 12.8
• PyTorch 2.11.0+cu128

Credits

• Stable Audio Open 1.0 — Stability AI
• opus-mt-mul-en — Helsinki-NLP / Tatoeba Translation Challenge
• NLLB-200 — Meta AI / FAIR
• Gradio, diffusers, transformers, accelerate — HuggingFace
• Wrapper development — Oscar Lumiere
• Assistance — Claude Code (Anthropic)