A voice agent should know when to listen.
We are developing conversational voice systems and investigating how they handle interruptions, hesitation and the transfer of a conversation to a human.
Our language work starts with European Portuguese (pt-PT). We welcome collaborators interested in evaluating real conversational behaviour, including interaction with people speaking a language they are learning.
Evaluating Interruption Handling in Voice Agents for European Portuguese and English
A proposed method for testing explicit stop requests, corrections, backchannels, hesitation and background speech. It defines measurements at the listener's audio output, so the whole system is assessed rather than a model's internal event timestamps.
Status: protocol published for discussion. No benchmark results or model rankings are reported in this version.
Read the protocol Download PDFWhat we want to learn
- How reliably does an agent yield when a person asks it to stop or corrects a detail?
- Can it distinguish an interruption from a short acknowledgement such as “uh-huh” or “pois”?
- Does a pause by a language learner cause the agent to take the floor too soon?
- How do playback buffering, background speech and the communication channel affect the experience?
Work with us
We welcome Portuguese-speaking reviewers, researchers working on spoken dialogue and teams willing to evaluate their voice systems using a shared protocol. Potential contributions include reviewing pt-PT test prompts, agreeing on annotation rules, running comparable tests and analysing failures.
Research contact: Valerii Kondrashov · valerii@goippro.com
GoIPpro, LLC · Delaware, USA
Our Behind the Glass journal records engineering experience. The research section complements that journal with explicit methods, limitations and, when available, reproducible measurements.