Start the demo and say a command in English
Your audio never leaves this page. The model listens for 14 English commands spoken on their own — it's deliberately tuned to ignore the same words inside a sentence or longer phrase, and to ignore all other speech. Watch each command's confidence rise live as you speak.
Measure the real-time factor on this device
RTF (real-time factor) is processing time divided by audio duration — an RTF of 0.01 means one second of audio takes 10 ms to process, so keeping up with live speech uses about 1% of one CPU core. This benchmark generates 60 seconds of audio right in your browser, pushes it through the keyword-spotting engine as fast as your CPU allows, and reports the measured RTF.
For reference, VoxRT's published browser number on this same model: 0.97% RTF on a 10-core desktop x86 (Chrome) — 103× realtime, 0.31 ms per 100 ms chunk. Mobile numbers land with the next release.
How it works: the ~190 KB WebAssembly runtime and ~1.28 MB model load from a CDN, then all inference runs on your CPU in this tab — a cache-aware streaming Conformer (636 K parameters) scoring all 14 commands 25 times a second. Detection quality on the speaker-disjoint held-out set: 0.9966 accuracy, 0.967 macro-F1. A command fires when its score stays above the threshold for 3 consecutive frames (~120 ms), then observes a 1 s per-command cooldown. The vocabulary is fixed — this published model ships with the SDK and is free for commercial use; custom command vocabularies are built per customer as a paid engagement. Requirements: a browser with WebAssembly SIMD128 — Chrome/Edge 91+, Firefox 89+, Safari 16.4+ — and microphone permission. Lower the threshold for easier triggering (more false positives); raise it for stricter matching. Built from the open @voxrt/kws-browser package on npm.
Usually paired with a wake word
In a shipping product, a command vocabulary rarely listens on its own. An always-on custom wake word runs first — the low-power trigger that brings the device to attention. Only once it fires does keyword spotting take over, for the few seconds a command is actually expected: "play", "pause", "next", "off".
See the pairing in a product context in the hands-free voice control and voice control for devices use cases.