Neuphonic open sources NeuDecide, a 43 MB audio-to-tool model with WASM browser demo

Neuphonic has open sourced NeuDecide under the Apache 2.0 license, a model that converts audio directly into tool calls with arguments, bypassing the traditional speech-to-text transcription step. The model files total 43 MB and are designed to run on a single CPU thread, making them suitable for resource-constrained environments. On the SLURP benchmark's tool-only task with 10 tools available, NeuDecide achieves 72.4% tool accuracy, which is approximately 3 times higher than Nvidia Parakeet + Google FunctionGemma (24.4%) and 3.5 times higher than Cactus (Whistle + Needle) (20.7%). Inference metrics on a single CPU thread show a time-to-call of 82 ms on a Samsung S24+ and 207 ms on a Raspberry Pi 5. The architecture consists of three ONNX components: an audio encoder, a tool encoder that integrates tokenized JSON tool definitions, and a decoder that generates tool calls token by token. Because tool lists are provided at inference time, no retraining is needed to change available actions. Neuphonic provides a WASM browser demo for testing presets like robot vacuums and car controls, as well as a Python package for local inference.