README.md

metadata

title: NeuDecide
emoji: 🎙️
colorFrom: blue
colorTo: yellow
sdk: gradio
sdk_version: 6.29.1
app_file: app.py
short_description: Speech in, tool calls out — no transcript in between
python_version: '3.12'

A demo of NeuDecide, a tiny (~55M parameter, 43 MB) voice-action model from Neuphonic. Record a spoken command, give it a list of tools as JSON, and it emits the matching tool call — no transcript in between and no GPU required.

"clean the bathroom" + [cleanRoom, goToBase, getBattery] → [{"name": "cleanRoom", "arguments": {"room": "bathroom"}}]

• Inputs: a short voice recording (≤ 30 s, English) and a JSON list of tool definitions (name, description, JSON-Schema parameters).
• Output: the tool calls the utterance maps to — a JSON list of {"name", "arguments"} dicts, or [] when no tool applies.
• How: three q4 ONNX graphs (audio encoder, tool encoder, decoder) run on CPU via ONNX Runtime with grammar-constrained decoding, using the neudecide package.

Tools are passed in at inference time, so you change what the model can do by editing the JSON — try the preset tool sets (robot vacuum, car control, smart home) or paste your own. Weights are bundled in model/; sample WAVs and preset tools are from the authors' repository (Apache-2.0).