# Interspeech 2026: MSU-Bench *A diagnostic benchmark for evaluating how well Large Audio-Language Models (LALMs) understand **who says what**, and **what happens between speakers**, in real multi-speaker conversations.* [](https://aslp-lab.github.io/msu-bench.github.io/) [](https://arxiv.org/abs/2606.22868) [](https://huggingface.co/datasets/ASLP-lab/MSU-Benchmark) [](LICENSE) [](#license) **Zhaokai Sun**\*, **Shuai Wang**\*, **Zhennan Lin**\*, Chengyou Wang, Dehui Gao, Yuang Cao, Chunjiang He, Pan Zhou, **Lei Xie**\*\* Audio, Speech and Language Processing Group (**ASLP@NPU**), School of Software, **Northwestern Polytechnical University**, China School of Intelligent Science and Technology, **Nanjing University**, China **Shenzhen Loop Area Institute**, China Base Model, **Li Auto**, China _{\* Equal contribution  ·  \*\* Corresponding author  ·  📮 `zksun@mail.nwpu.edu.cn`}

Figure 1. Two-tier task hierarchy of MSU-Bench, from speaker grounding to multi-speaker reasoning.

Figure 2. Capability definitions and representative tasks in MSU-Bench.

Figure 3. MSU-Bench construction pipeline: dialogue-quality assessment → speaker-aware annotation → speaker-referenced QA generation → human-in-the-loop quality control.

Figure 4. Full per-task results across the 16 MSU-Bench tasks (see paper for details).