Multi-speaker automatic speech recognition (MSASR) has advanced rapidly, particularly with the rise of end-to-end speech large language models. However, existing benchmarks do not adequately cover the full range of multi-speaker conditions and scenarios, leaving evaluation gaps.

To address this, the paper introduces MS-Bench, a condition-stratified benchmark designed to assess MSASR systems across a broader set of conditions. The benchmark aims to provide a more thorough and systematic evaluation than current options.

The work is available as an arXiv preprint. While the abstract does not detail the benchmark's composition or results, its motivation is clear: to close the gap in multi-speaker evaluation and support further progress in the field.