SmoothConv and DuplexConv: Complementary Mandarin Multi-Party Conversational Speech Corpora for Speech Interaction

Abstract

Recent advances in large audio language models (LALMs) have driven the development of natural and intelligent speech interaction systems. Such systems need to model complex conversational behaviors, including turn-taking, overlapping speech, and speaker coordination. Multi-party conversations provide a realistic setting for studying these behaviors, yet existing Mandarin conversational corpora often lack synchronized participant-level speech tracks and comprehensive annotations. In this work, we introduce SmoothConv and DuplexConv, two complementary Mandarin multi-party conversational speech corpora totaling 2,100 hours. SmoothConv provides human-verified conversations for reliable analysis and evaluation, while DuplexConv offers large-scale automatically annotated conversations through a scalable pipeline for model training. Both corpora provide synchronized participant-level speech tracks and multi-dimensional fine-grained annotations. We further release the SmoothConv Benchmark and evaluate these resources on speech separation, multi-speaker automatic speech recognition (MSASR), and turn detection tasks. Experimental results demonstrate the utility of the proposed resources for multi-party speech interaction modeling. The datasets, benchmark, and related resources are publicly available.