A voice assistant that holds the thread of a conversation

The problem

Voice assistants break in the places people actually use them. An accent or a noisy room defeats the recognition. Each request is treated as though the previous one never happened, so anything that needs a follow-up has to be re-stated in full. And a pause long enough to notice turns what should be a conversation into a form you happen to be filling in out loud.

What we built

• Speech recognition tuned for varied accents and background noise rather than for a quiet room
• A conversational engine that keeps context across multi-turn exchanges, so a follow-up can just be a follow-up
• Response generation over LLM models instead of a fixed set of intents
• Multilingual input, so people can speak in the language they think in
• An Android app designed around voice as the primary interface, not as an add-on to a keyboard

What made it hard

Four problems have to be solved simultaneously before any one of them counts: recognition that survives accents and noise, understanding that happens in real time, responses that are actually relevant to what was asked, and a mobile experience where talking to it is genuinely easier than typing. Get three right and the fourth still sinks it.

How we built it

• Fine-tuned the recognition pipeline against varied accents and noisy environments
• Carried conversational context across turns, so earlier parts of an exchange inform later ones
• Optimised the inference pipeline for sub-second responses — latency is what makes a conversation feel transactional
• Designed the voice interface for accessibility, so it works across technical comfort levels rather than assuming a power user

Who it is for

Android users who want assistance by speaking — hands busy, screen out of reach, or simply faster out loud — including people for whom typing is the barrier rather than the convenience.

See VoxaAI in the portfolio