Put voice inside the device — no cloud account per unit
Appliances, speakers, kiosks, robots: VoxRT runs the whole voice pipeline on the device's own CPU. No companion cloud subscription baked into your unit economics, no special silicon on the BOM, and no microphone audio leaving hardware that carries your logo.
Why device makers keep voice local
No recurring cost in the unit economics
Cloud voice means paying per device, per month, for the life of hardware you sold once. On-device voice is a fixed engineering cost — the published runtime and models carry no per-device fees.
Runs on the hardware you already spec'd
The runtime is validated on Cortex-A53-class boards — Raspberry Pi 3 territory. No DSP, no NPU, no BOM change; the application processor you already have is enough.
Immune to the internet
No outage takes the feature down, no deprecated cloud API bricks shipped units, no end-of-life announcement strands your customers. Voice works for the whole life of the device.
Start demo and say "Hey Assistant"
Your audio never leaves this page. The model highlights "Hey Assistant" while ignoring all other speech — including similar-sounding words and phrases like "Hey sister" or "Assist me".
"Hey Assistant"
Devices that answer for themselves
From the kitchen counter to the factory floor — voice that lives entirely in the firmware.
Home appliances"Stop" over a running blender
Range hoods, ovens, washers — command sets tuned for the appliance's own noise, working with wet or full hands.
Speakers & soundbarsA brand wake word + transport controls
The product answers to its own name and handles play/pause/volume locally — no cloud account per speaker sold.
Fitness equipmentMid-workout control
Treadmills and trainers that respond while both hands grip the bars — resistance, speed, pause, resume.
Kiosks & vendingWalk-up voice, no back end
Ordering and check-in kiosks that listen locally — nothing to provision, no per-kiosk cloud identity to manage.
Robots & toysAnswers its name, on-board
Consumer robots and interactive toys that respond even where Wi-Fi doesn't reach — and never stream a child's voice.
Building & facility controlsPanels in loud rooms
Machine-room and utility panels operated by voice where gloved hands and safety rules make touch impractical.
From spec sheet to talking product
Pick the interaction model
A wake word ("Hey Brand") to get the device's attention, a command vocabulary for control, or natural phrases mapped to intents — most devices combine a wake word with one of the other two.
We tune models for your hardware and acoustics
Your device's own noise, its microphone placement, its far-field reality — models are trained and thresholds tuned against the environment the product actually ships into.
Integrate one C library
A single ~480 KB shared library on Linux aarch64 with Python, Node.js, Go, C/C++ and Rust wrappers — or the native iOS/Android SDKs for app-connected hardware.
Ship it and forget it
No cloud keys to rotate, no per-device provisioning, no monthly invoice tied to units in the field. The voice stack is part of the firmware, full stop.
The numbers behind the promise
Runtime numbers are measured on real boards and published — see the benchmark methodology. The browser demo below runs the same wake-word detector your firmware would.
Pipeline. Microphone → on-device VAD (the always-on gate; ~1 ms per 32 ms frame) → wake word and/or keyword spotter → events straight into your firmware or application process. Everything rides one runtime and one binary.
Platforms. iOS 16+ (Swift Package Manager) and Android 8.0+/API 26 (JitPack/Gradle), plus Linux aarch64 for device-class hardware — validated on Raspberry Pi 3/4/5 and Pi Zero 2 W, NVIDIA Jetson Nano and Orin Nano, AWS Graviton and similar Cortex-A53/A55 boards — shipped as a single ~480 KB shared library with Python, Node.js, Go, C/C++ and Rust wrappers. The Linux quickstart is in the docs: wake word on Linux.
Licensing. The runtime and published models are free for commercial use, including in hardware you sell, with no per-device fees. Custom wake phrases and command sets are one-time tuning engagements — see licensing.
Voice in hardware, answered
Do we need a special voice chip, DSP, or NPU?
No — the runtime is CPU-only and validated on Cortex-A53-class boards (Raspberry Pi 3 and up). If your device has an application processor from the last decade, it can likely run the pipeline.
Is there a per-device license fee?
No. The published runtime and models are free for commercial use in shipped hardware, with no per-unit royalties. Custom models — your brand's wake phrase, your appliance's command set — are one-time paid engagements. See licensing.
What happens when the device is offline?
Nothing changes. The wake word, commands, and intent handling are all local, so the voice interface works identically with no connectivity — including for the entire life of the product after any companion cloud is gone.
Can the device have its own wake word?
Yes — that's the most common request. See the branded wake word use case for what the engagement looks like end to end.