VoxNexus just launched the most ambitious pitch in voice AI this year: a phone that hears everything, understands dozens of languages, and uploads absolutely nothing.
The Singapore-based startup, led by CEO Charles Shi, announced two products on October 23 — VoxEdge and VoxSense — both unveiled publicly at the MediaTek Dimensity 9600 launch in Madrid on October 19. The idea is simple and radical at once: put the whole voice pipeline on the device itself.
Two products, zero cloud
VoxEdge is an on-device voice stack for Android: speech recognition in 36 languages, translation in 103, and speech synthesis in 45 — all offline, with no connection, no data bundle, and, the company says, no cloud token bills. Every query is free once the model is installed. VoxNexus claims its speech recognition model packs just 65 million parameters yet posts a lower word error rate than open-source models nearly ten times its size.
Then there's VoxSense — the attention-grabber. It's on-device ambient listening: 12 hours of continuous listening on 162 mAh, roughly 2.5% of a typical 6,500 mAh phone battery, with all processing on the device and no audio ever sent to the cloud.
Why on-device voice matters now
Across much of Africa and South Asia, voice AI has been held back by scarce training data, dialect variation, and everyday code-switching — plus the cost of cloud inference at scale. "Under-resource languages aren't small languages. Hundreds of millions of people speak them. Until now, AI models just haven't learned them yet," Shi said. VoxNexus is aiming squarely at that gap, and it's shipping to developers and device makers through a license plus SDK, with a free trial at voxnexus.ai.
The honest caveats
This is PR-wire coverage, so the performance claims are VoxNexus's own — no independent benchmarks yet, and the company frames them as internal evaluations. And a phone that listens all day will live or die on user trust, no matter how airtight the on-device promise sounds. For today, cloud-based AI transcription tools remain the practical default for most people.
The take
Voice agents are having their moment — Mistral's Voxtral 3 stack and ElevenLabs' real-time models landed just days apart — and VoxNexus is betting the next frontier isn't smarter voices, but voices that never leave your pocket. If the battery and accuracy claims hold up, that's a genuinely new category.
Comments
Post a Comment