Woosh Studio — Web UI

A web UI, REST API, and MCP server for Sony Research's Woosh sound-effect foundation model, packaged for GPU-accelerated deployment on an NVIDIA Jetson Orin.

Type a prompt — laser firing, glass shattering, thunderstorm rain — and generate sound effects with either the fast distilled DFlow model or the full-quality Flow model.

Features

• Two models — DFlow (distilled, ~4 steps, sub-second on GPU) and Flow (full quality, adaptive ODE solver).
• Batch generation — up to 4 variations per prompt, shown as a thumbnail grid with inline players, filenames, sizes, and download links.
• Export settings — OGG / MP3 / WAV at a selectable sample rate (8–96 kHz) and bitrate; the download links are encoded to match.
• Live progress that tracks the sampler's steps (DFlow steps / Flow solver evaluations).
• REST API (Gradio) with copy-paste CLI / JavaScript / Python snippets in the UI.
• MCP server exposing a generate_sound tool for MCP clients (Claude Code, Claude Desktop, …).

Running

The app runs as a systemd --user service (woosh-ui.service) that auto-starts on boot and auto-restarts on failure:

systemctl --user status  woosh-ui.service
systemctl --user restart woosh-ui.service
journalctl --user -u woosh-ui.service -f

It serves on port 7860 — open http://localhost:7860 (or http://<host>:7860 from another machine on your network, where <host> is the server's hostname or IP).

To run it manually instead:

uv run python woosh_ui.py --server-name 0.0.0.0 --server-port 7860

The Woosh model checkpoints are not included in this repo. Place them under checkpoints/ (checkpoints/Woosh-DFlow, checkpoints/Woosh-Flow) — see the upstream repo for download links.

MCP

The MCP server (SSE transport) is exposed at http://<host>:7860/gradio_api/mcp/sse (replace <host> with the server's hostname or IP).

Add it to an MCP client — e.g. Claude Code:

claude mcp add --transport sse woosh http://<host>:7860/gradio_api/mcp/sse

…or in .mcp.json / claude_desktop_config.json:

{
  "mcpServers": {
    "woosh": {
      "type": "sse",
      "url": "http://<host>:7860/gradio_api/mcp/sse"
    }
  }
}

Tool: generate_sound(prompt, model="DFlow", batch_size=1, cfg_scale=4.5, seed=-1, output_format="ogg", sample_rate=48000, bitrate="256k") → list of audio file URLs.

GPU build (Jetson Orin)

This deployment targets JetPack 6 / CUDA 12.6 / Python 3.10. torch, torchaudio, and torchvision are sourced from the Jetson AI Lab index (see pyproject.toml), and numpy is pinned <2 to match the Jetson torch ABI. uv sync reproduces the CUDA-enabled environment:

uv sync
python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
# -> 2.8.0 12.6 True

Credits

Built on SonyResearch/Woosh (Apache-2.0). Model weights are © their respective authors and are not redistributed here.