SQuTR Benchmark Tests Spoken Query Retrieval Under Noise

Researchers have introduced SQuTR, a robustness benchmark for spoken query to text retrieval systems. The benchmark includes a large-scale dataset of 37,317 unique queries from six English and Chinese text retrieval datasets, covering multiple domains. Speech is synthesized from 200 real speakers, and 17 categories of real-world environmental noise are mixed under controlled SNR levels. This allows reproducible evaluation from quiet to highly noisy conditions. The work aims to address the limitations of existing evaluation datasets, which are often restricted to simple queries under constrained noise conditions.