A multidialect multidomain Tibetan speech dataset for speech and language processing

Tibetan is a morphologically complex and relatively low-resource language family, widely spoken across the Tibetan Plateau and characterized by substantial dialectal variation. To address the limited availability of high-quality open speech resources, we construct and release a multi-dialect, multi-domain Tibetan speech corpus that covers three major dialects (Ü-Tsang, Amdo, and Kham) and three application domains (general, medical, and financial). The corpus contains approximately 20.39 hours of speech from 13 native speakers, accompanied by manually verified text transcriptions and detailed demographic metadata. To assess the usability and reliability of the dataset, we conduct baseline experiments on dialect identification and domain classification. The results demonstrate stable acoustic feature distributions across dialects and consistent cross-domain modeling performance, supporting the internal consistency of the dataset and the reproducibility of the reported baselines. This corpus fills a critical gap in publicly available Tibetan speech resources and provides an important foundation for research on low-resource speech technologies and the preservation of linguistic diversity.