Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

Breaking01 / AI

AVCC adds human benchmark for next-speaker prediction

The benchmark compares human and model predictions on the same conversation samples recorded with a single camera and microphone.

A researcher seen from behind compares blurred source passages, an external glossary, and contrasting AI translation outputs on two monitors in a language technology lab. AI-generated image

AVCC researchers published a next-speaker prediction benchmark on 11 Nov using 31 hours of conversation recorded with a single camera and microphone, according to their paper.

The abstract specifies that the benchmark evaluates whether the same speaker continues or another speaker takes over after a gap when everyone is silent. It includes 865 cases of a speaker continuing and 865 cases of a speaker change.

The researchers said they added human predictions collected on the same evaluation cases as a comparison baseline. They also released transcripts with word-level timing for the full conversation corpus and individual human responses.

The abstract names the audio-only RoleVAP and multimodal MuVAP as the models evaluated; multimodal models use more than one type of information. The researchers said they also released scripts to reproduce the published figures.

According to the abstract, earlier next-speaker prediction research used conversation recordings with separate sensors for each speaker. Studies also differed in how they defined the events to evaluate.

The abstract does not disclose prediction scores for each model or the performance gap between models and people.