An operator needed to speak a command over a handheld radio and have an uncrewed system act on it, in a tactical setting with no cloud and no room for a wrong action, in degraded, GNSS-denied, offline conditions. The command set includes irreversible, lethal-capable verbs, so a false positive is a safety event rather than a user-experience annoyance. The obvious build is speech to text then parse the text, which drags a large, latency-heavy, error-prone transcription stack into a place that can afford none of the three.
The stated need was accurate voice control. The real requirement sat one layer below the words. Because the verbs are lethal-capable, the governing requirement is not recognition accuracy. It is that the default behavior under any uncertainty is to do nothing. That reframe, from recognize speech well to fail closed by construction, is what turned a model problem into a systems problem.
The reflex, and the fix.
Speech to text, then parse the text
The standard architecture, with mature tooling and a fast path to a demonstration.
It drags a large, latency-heavy, error-prone transcription stack into an air-gapped edge module that can afford none of the three, and it puts an unbounded intermediate representation inside a safety-critical path.
Map speech directly to a bounded intent set
Building direct audio-to-intent recognition without transcription tooling, and accepting a deliberately bounded verb set.
The largest and most fragile subsystem removed in one decision, an enumerable output space, and a safety argument that can be made.
Skip transcription entirely, because a finite output space is the precondition for every safety property downstream of it.
Silicon and edge, one module, no cloud
The whole chain runs on a single small edge module, air-gapped, with no telemetry home and no remote correction once fielded.
Metal and ingest, the real radio path
Audio arrives through a hardware interface with real push-to-talk keying. This is the metal end most voice systems avoid by testing on clean audio, and it is where the accuracy that matters is measured.
Front end, disciplined and separated
Voice-activity detection with disciplined resampling, and strict separation between training and evaluation so the reported number means something.
Model, compact and frozen
A compact adapted speech encoder, roughly 25 million parameters, frozen at deploy, producing intent directly across a bounded verb set. No transcription anywhere in the path.
Safety, the gate that defines the program
An intent and safety gate where any uncertainty routes to a HOLD and no action is taken. Not a tunable threshold: the default under uncertainty is doing nothing, by construction.
Actuation, protocol translation out
Translation to the flight controller and the tactical picture, with the gate enforced beneath the layer most likely to be wrong.
A single headline number hides where a system fails. This work was scored on the dimensions that actually decide whether it holds in production, measured on real, held-out cases rather than the demo path.
Latency and accuracy are measured. The second radio path is modelled with hardware pending and is labelled. Disclosure tier confirmed with the founder.
A transcription pipeline that demonstrates well on clean audio indoors, cannot state what it does when uncertain, does not fit the power envelope, and cannot be certified because nobody can enumerate what it might conclude.
Audio in to command out in 13.4 ms against a 22 ms budget, measured. Command accuracy of 0.862 through a real radio path, measured, and 95.7% on-device under clean conditions. A second radio path sits at 0.846, modelled, with hardware pending, and stays labelled as such.
What it owns, and what it hands to a person.
The corpus was single-speaker, so speaker independence is named here as an open workstream rather than hidden. The two accuracy figures measure different things and are reported that way: 95.7% is on-device under clean conditions, 0.862 is through a real radio path, which is the harder and more honest number. A figure you cannot defend in a technical conversation is worse than no figure. The bounded verb set is a deliberate constraint and the source of the safety properties, so extending it is real work rather than configuration.
The deciding decisions were not machine learning decisions. Removing transcription was architecture. Enforcing the hold beneath the actuation layer is systems design. Validating against real firmware is firmware work. The recognition model, the part a specialist would have been hired to build, is downstream of all three, and building it well would not have made the system fieldable. The discipline carries to any system where an incorrect action is not recoverable: keep the output space enumerable, make the safe behavior structural rather than configured, and enforce it beneath the layer most likely to be wrong.