University of Washington's hearing assistant finds your conversation partners by tracking turn-taking

At a glance

The short version: it tracks who takes turns with you to find your partners; it's a lab prototype.

Do this first: watch for a named earbud or hearing-aid maker piloting it, and an independent live trial.

University of Washington researchers have built headphones that work out who the wearer is talking to without being told. The wearer's own voice is the anchor: the system tracks who takes turns with it and tunes out voices that overlap too often. It needs two to four seconds of audio to find partners, copes with one to four of them, and was presented at the EMNLP conference in Suzhou on 7 November 2025. The team says it could one day reach hearing aids, earbuds and smart glasses.[1] [2] [3]

Two models run at different speeds. A fast one handles each 12.5-millisecond slice of audio, keeping the conversation and dropping the rest. A slower one updates once a second, studies who spoke when, and hands the fast model a summary of the conversation's rhythm. IEEE Spectrum's Meghie Rodrigues explains why: natural-sounding audio must be processed in under 10 milliseconds, yet spotting turn-taking takes one to two seconds of context. Without the slow model, the paper's reported gain falls from 12.48 dB to 1.45.[3] [1] [2]

The fast half has to live on the headphones, the paper argues, because a Bluetooth link adds 10 to 30 milliseconds of delay and the internet 100 to 200. In testing it ran on an embedded Orange Pi 5B, averaging 8.9 milliseconds per 12.5-millisecond chunk, while the slow model, on an Apple M2 chip, needed 41.3 milliseconds per second of audio. ASME's Poornima Apte, writing in June 2026, describes that larger model as living on a connected smartphone, and the university's Shyam Gollakota expects to shrink the whole system onto a tiny chip in an earbud or hearing aid.[3] [4] [1]

In the listening study, 11 people scored processed clips 4.30 out of 5 against 1.88 for the originals, each hearing six recorded conversations from the Candor dataset. On the team's recordings of real conversations, accuracy was 85.0% for two-person chats and 73.4% for three-person ones, with speech-quality gains of 7.84 and 6.00 dB; IEEE Spectrum's up-to-14.6 dB figure comes from the paper's synthetic mixes.[3] [2] [1]

Te-Won Lee, who has recently developed similar technology for commercial use, told IEEE Spectrum that he sees the prototype's very low latency as a major strength, though most environments don't offer four people neatly taking turns. The paper reports the same weak spot: partner quality fell below zero when the wearer stayed silent for over two minutes, then recovered once they spoke. Its authors suggest a physical button to switch the assistant off.[2] [3]

Three of the paper's four authors list Hearvana AI, a startup that announced a $6 million pre-seed round, an early funding stage, in November 2025 and says it plans commercial pilots with consumer-electronics partners. The code is open source. Watch for two things: a named earbud or hearing-aid maker piloting the approach, and an independent trial in which people wear it live through a dinner full of interruptions and music.[3] [5] [1]