Learning from use
Proposes a governed learning loop that turns user corrections and usage signals into improvements across knowledge, search, instructions, memory and model training, with evaluation and human approval before release.
Read the paperRESEARCH / ENTERPRISE AI
Voice and screen, working together
An architecture for multilingual voice assistants that coordinate spoken answers with on-screen information, using native audio models, isolated sessions and a compact knowledge catalog to support responsive public-kiosk interactions.
Read on this pageAt a glance
A public-facing voice assistant needs to respond naturally, handle language changes and show the right information on screen at the same time. This technical report describes a native-audio architecture for that setting, using a single model for the spoken interaction and coordinated screen updates.
The design isolates each session, uses a compact knowledge catalogue and connects visual changes to the same actions that produce the spoken response. The report explains its latency budget and the practical failures encountered in preview-model deployments, including the fixes used to address them. It is relevant to retail, service and information kiosks where the experience depends on the whole conversation, rather than the quality of a transcript or a standalone text answer.
What the paper covers
Native audio keeps the spoken interaction within a single model contract.
Session isolation and coordinated tool actions connect voice answers to screen state.
Latency is measured across the customer experience rather than one processing component.
Research context
The report describes the documented deployment and its measured response behaviour. Long sessions beyond the model's limit, live inventory refresh and multi-host scaling are outside its current scope. Reported timing should be validated under each deployment's models, network and workload.
THE ORIGINAL WHITEPAPER
Download PDF Read it here.
The complete paper, with its original figures, tables and references.
29 pages · Read without leaving this pageFROM RESEARCH TO YOUR WORK
Discuss this research Bring your questions.
Discuss the approach, evidence and relevance to your team with the people building it.