What this is
Neuphonic, a London voice AI startup, this week open-sourced a model called NeuDecide under the Apache 2.0 license. It weighs just 43MB. It does something counterintuitive: when a user speaks, the model directly outputs instructions specifying which tool to call and what parameters to pass—skipping the traditional step of converting speech to text first.
On SLURP (a public speech tool-calling benchmark), it picks the correct tool out of 10 candidates with 72.4% accuracy. By comparison, chaining Nvidia's Parakeet (a speech-to-text model) with Google's FunctionGemma (a text-to-function-calling model) into a two-step pipeline only hits 24.4% accuracy—NeuDecide is 3x better.
The hardware bar is surprisingly low: 46ms from speech to command on a MacBook M3, 82ms on a Samsung S24+, and it even runs on a Raspberry Pi 5—that sub-$100 mini computer (206ms).
Industry view
Supporters see this as a landmark moment for on-device AI (models running locally on the device, not relying on the cloud). Cloud-based large models are expensive, slow, and require uploading voice data. NeuDecide lets robots, smart home devices, and in-car systems understand and execute commands even without internet.
But the skepticism deserves a hearing: that 72.4% accuracy was measured on a relatively clean benchmark with only 10 tools. In the real world, accents, background noise, and multi-turn dialogue will all drag accuracy down. And when voice commands execute incorrectly—"dim the living room lights" understood as "turn off the living room lights"—the cost isn't trivial.
Another contrast worth noting: when giants like Google and Nvidia stack their strongest models together, they still lose by 3x. On specific tasks, specialized small models have caught up with—or surpassed—the patchwork approach of bolting general-purpose large models together.
Impact on regular people
For enterprise IT: Customer service systems, in-car platforms, and smart home vendors can deploy voice understanding locally—no need to upload user audio to the cloud, benefiting both privacy and bandwidth costs.
For working professionals: "Speak to pull up a spreadsheet or write into a document" could become the next productivity entry point, but that requires product-level integration. It won't happen automatically in the short term.
For the consumer market: Smart speakers and in-car voice systems will get better at "understanding what people actually mean," but 72% is still a long way from 99%—don't overestimate how fast this replaces touchscreens anytime soon.