Duplex-MPE benchmarks multi-party full-duplex speech: MiniCPM-o 4.5 leads on three of four scores

PKU · hf · 2026-09-29

Duplex-MPE tests when a full-duplex speech assistant should answer, stay silent, or stop speaking in multi-party conversations. The benchmark has 2,000 scenarios with 3-4 human speakers plus an assistant, paired across explicit/implicit addressing, with continuous audio and no transcripts. Among five open-weight systems evaluated, MiniCPM-o 4.5 leads on three scored capabilities; frequent speech from others coexists with wrong answers or failure to stay silent. A transcript-based Gemini 3.1 Pro reference responds 64.3 points more to explicit than implicit requests, while speech systems show no significant difference — current models can't tell if they're being addressed.

Original post →

More from Models

Models channel →