MultiTalkBench

MultiTalkBench MultiTalkBench is the first benchmark to jointly evaluate long, multi-party, and bilingual full-duplex dialogue. It tests speech-to-speech systems on: (a) Long interactions — conversations longer than ten minutes, with explicit probes for long-range entity tracking and topic coherence. (b) One-model-many-user multi-party interaction — quantitative addressee-selection and turn-taking metrics. (c) Chinese–English bilingual ability. To our knowledge, no prior benchmark… See the full description on the dataset page:

Source: Hugging Face Hub (MultiTalk/MultiTalkBench). Metadata imported from the dataset’s Hub tags.