Research on End-to-End Multilingual Speech-to-Tibetan Text Translation System

With the rapid development of multilingual speech translation technology, cross-linguistic communication has gradually become an important demand in the era of globalization, especially for the translation of minority languages, where technological breakthroughs are particularly crucial [4]. This paper combines the Conformer speech encoder and the DA-Transformer language decoder to achieve translation from Mandarin and English speech to Tibetan text. The model adopts a non-autoregressive architecture, which directly converts speech signals into target text without relying on traditional cascade models, thus avoiding error accumulation and processing delays that may occur in cascade systems [1]. Due to the relative scarcity of Tibetan corpora, especially the lack of high-quality aligned speech-to-text translation data, this paper collects and processes datasets for Mandarin, English, and Tibetan [8]. Experimental results show that the proposed system outperforms the baseline model in translation quality and decoding speed.