New Shared 'File Format' Could Make AI Voice Assistants Less Confused

When AI voice tools work together — one listening, one taking notes, another answering questions — they usually just pass text to each other. Everything else gets dropped: who was speaking, exactly when they said it, how they sounded, and whether the information was guessed or heard directly. Worse, each tool packages those details in its own private format, so when one AI hands off to the next, the handoff is fragile. Think of contractors working from different blueprints for the same house.

OneVoice, from researchers Vipul Charugundla, Dancheng Liu, and Jinjun Xiong, is a fix for that handoff. It's a lightweight, shared file format — the researchers describe it as an intermediate representation, meaning a common middle layer everything can read. It stores speech information as structured records with stable IDs, layered transcripts, explicit timing links, and provenance (a record of where each detail came from). It's built on JSON, a plain, widely used data format, so it's easy for many tools to read.

The team tested it in two multi-agent workflows — one aggregating events, one linking sound details to exact moments in time — using three different language models. Compared with letting each AI invent its own handoff format, OneVoice substantially improved how reliably information was combined and how well timing relationships survived the trip. The paper was posted on arXiv, a preprint site, meaning it hasn't been formally peer-reviewed yet.

Here's the honest catch: this is plumbing, not a product. There's no app, no release date, and no known company using it. The tests were lab-sized, not real-world call centers or meeting rooms. But the underlying idea matters — the messy glue between AI tools is often where they fail. If a standard like this catches on, voice AI that books appointments, transcribes meetings, or summarizes calls could become noticeably more trustworthy, because it would stop losing the details that make a conversation make sense.