An operator speaks a command into a handheld radio and an uncrewed system acts on it. The conditions are degraded, GNSS-denied and offline. There is no cloud to lean on and nobody can correct the system once it is fielded. Some of the verbs it understands are irreversible and lethal-capable, so a command it hears that was never given is a safety event, not an inconvenience. The usual way to build voice control, transcribe the speech and then parse the text, brings a large, slow and error-prone stack into a place that can afford none of those three.
A voice command system can be built to choose from a list, and a list can be checked.
Most people picture voice control as two machines in a row. The first turns speech into text, the way dictation does. The second reads the text and works out what was meant. The first machine is enormous, because it has to be ready for any word in the language, and its output is open-ended: any sentence at all can come out of it. Everything downstream then has to cope with that openness, including sentences nobody imagined.
There is a second way. Sound is cut into short overlapping frames, each frame is turned into a picture of how its energy is spread across frequencies, and a compact neural encoder reads that sequence and condenses it into a summary of what was said. A small classifier on top picks one entry from a fixed list of commands, plus an entry that means none of these. No words are ever written down. The system cannot produce a command that is not on the list, because there is nowhere for one to come from, and every entry on the list can be tested on its own.
A handheld radio adds its own physics. It passes only a narrow band of frequencies, may squeeze the voice through a codec, clips loud syllables, and adds a click and a burst of noise each time the transmit key goes down or up. A model that learned from clean audio meets a different signal in use. That is why accuracy on clean audio and accuracy through the radio are separate numbers answering separate questions, and why they are never blended into one.
The finite list is also what makes fail-closed design possible. A fail-closed system is one whose resting state is the safe state, and which needs positive evidence to leave it. A railway signal that falls to red when its wire breaks works this way, and so does a machine guard that stops a press when it is opened. For voice, it means silence, noise, a clipped transmission, a word outside the list, an ambiguous result or an internal fault all lead to the same outcome: hold, and do nothing. Only a confident match to a known command moves anything. Accuracy still matters, because a system that holds too often is useless. The fail-closed structure is what makes accuracy a question of usefulness and keeps it from becoming a question of safety.
Accurate voice control was the brief, and accuracy turned out to be the second question. When some of the verbs cannot be taken back, the first question is what the system does when it is unsure, and the only acceptable answer is nothing. Recognising speech well is a property of a model. Doing nothing under any uncertainty, by construction, is a property of the whole system. Designing for that second property is what turned a model problem into a systems problem.
Voice systems fail at the edges of their input, and a command system has to decide what the edges mean.
Speech recognisers are usually judged on clean, in-vocabulary audio from cooperative speakers, and on that material they do well. Their trouble is at the margins: a word that is nearly a command, a transmission cut off halfway, a second voice in the background, a speaker the model has not heard before. A dictation tool that mishears produces a typo. A command system that mishears produces an action.
The second pattern is treating safety as a threshold. Teams add a confidence cutoff and call the system safe, but a cutoff is a number someone can change, and models can be confidently wrong. When safe behaviour depends on a tuned value inside the model layer, the layer most likely to be wrong is also the one trusted to catch its own mistakes.
The reflex, and the fix.
Speech to text, then parse the text
It is the standard architecture. The tooling is mature and it reaches a convincing demonstration quickly.
It brings a large, slow, error-prone transcription stack into an air-gapped edge module that can carry none of that weight. It also places free text, an output nobody can list in advance, in the middle of a path that has to be argued safe.
Map speech directly to a bounded intent set
Direct audio-to-intent recognition built without transcription tooling, and a verb set kept deliberately small and closed.
The largest and most fragile subsystem gone in one decision, an output space that can be listed in full, and a safety case that can actually be made.
Leave transcription out entirely. A finite set of possible outputs is the precondition for every safety property further down the chain.
One edge module, no cloud
The whole chain runs on a single small module. It is air-gapped, it sends nothing home, and once fielded nobody can reach it to correct it. What it does on the bench is what it will do in use.
The real radio path, push-to-talk and all
Audio arrives through a hardware interface with genuine push-to-talk keying. This is the messy end of the problem that most voice systems sidestep by testing on clean recordings, and it is where the accuracy figure that matters is measured.
A disciplined front end
Voice-activity detection decides where speech starts and stops, and resampling is handled carefully so the model hears audio shaped like the audio it learned from. Training and evaluation material are kept strictly apart, so the reported number describes speech the model had never heard.
A compact model, frozen at deploy
An adapted speech encoder of roughly 25 million parameters maps audio straight to one of a bounded set of intents. Its weights are fixed when it ships, so the model that was evaluated is the model in use. Nowhere in the path is speech turned into text.
The gate that defines the programme
Every candidate intent passes an intent and safety gate. Anything short of a confident, well-formed command routes to HOLD, and HOLD means no action. This is how the system is built, so there is no setting at which doubt produces movement and no dial anyone can turn to create one.
Translation out, with the gate underneath
Accepted commands are translated into the protocol the flight controller expects. The hold is enforced beneath the recognition model, the layer most likely to be wrong, so no output of that model can reach actuation without passing it.
What each choice cost, and what it bought.
ALeaving transcription out
Dropping speech-to-text gave up the most mature tooling in the field and the quickest route to a demonstration. Every model component had to be built for direct intent recognition. What it bought was larger than the latency and power savings. With no free text in the middle, the full set of things the system can conclude is a short list, and a short list can be tested entry by entry and defended in a safety review. A transcription path would have been quicker to show and impossible to bound.
BMaking the hold structural
The simplest form of caution is a confidence threshold inside the model, tuned until the demonstration looks right. Here the default under uncertainty is HOLD by construction, enforced beneath the recognition model so that no model output can get around it. The cost is a system that sometimes declines a command an operator really gave, and has to be asked again. The gain is that no setting, no tuning mistake and no confident model error can turn doubt into movement.
CMeasuring through the radio
Accuracy on clean audio is easy to obtain and flattering. Measuring through a real radio path with real push-to-talk keying meant building the hardware interface first and accepting a much harder test. Both results are reported, each under its own conditions: 95.7% on-device under clean conditions, and 0.862 through a real radio path. Each answers its own question, and only the second describes what an operator will experience.
DFreezing the model and bounding the verbs
A model that keeps learning in the field and a verb set that grows by configuration both sound flexible, and both would undermine the evidence. Freezing the weights at deploy means the model that was evaluated is the model in use. Keeping the verb set deliberately small means extending it is engineering work with its own validation. The system is slower to change as a result, and that slowness is the reason anything about it can be stated with confidence.
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 as modelled wherever it appears.
A transcription pipeline that demonstrates well indoors on clean audio. It cannot say what it does when unsure, it does not fit the power envelope, and it cannot be certified, because nobody can list everything it might conclude.
Audio in to command out in 13.4 ms against a 22 ms budget, measured. Command accuracy of 95.7% on-device under clean conditions, measured. Command accuracy of 0.862 through a real radio path, measured, and reported as its own figure because it answers a different question. A second radio path is modelled while its hardware is pending, and it carries that label wherever it appears.
What it owns, and what it hands to a person.
The training corpus came from a single speaker, so speaker independence is an open workstream and is named as one. The two accuracy figures measure different things and are kept apart. 95.7% is on-device under clean conditions. 0.862 is through a real radio path, which is the harder and more honest test. A figure that cannot be defended in a technical conversation is worse than no figure. The bounded verb set is deliberate and is the source of the safety properties, so extending it is engineering work with its own validation, never a configuration change.
The decisions that settled this system were not machine learning decisions. Removing transcription was architecture. Enforcing the hold beneath the actuation layer was systems design. Validating against real firmware was firmware work. The recognition model, the part a specialist would have been hired to build, sits downstream of all three, and building it well would not on its own have made the system fieldable. The same discipline applies wherever a wrong action cannot be undone: keep the output space enumerable, make the safe behaviour structural, and enforce it beneath the layer most likely to be wrong.
If an action cannot be undone, decide what the system does when it is unsure before deciding how accurate it must be, and make that default doing nothing.
A system that chooses from a fixed list can be tested exhaustively and argued safe. A system with free text in the middle of a control path cannot.
Safety that lives in a tunable threshold inside a model can be tuned away. Enforce it in a separate layer that every model output must pass through.
Measure accuracy through the channel the system will actually use. A figure from clean audio describes a different system.
Report each figure with its conditions attached, and keep figures from different conditions apart. Blending them hides the one that matters.