Tibetan Dialect Conversion Without Using Parallel Corpus

In this study, we propose a two-pass pipeline for Tibetan Dialect Conversion systems. To address the challenges of limited training data and complex morphological structures of Tibetan script, we adopt the Byte Pair Encoding (BPE) algorithm to construct a subword vocabulary. For the ASR component, we leverage the w2v-BERT 2.0 model, which combines self-supervised speech representation learning with contextualized language modeling. For the TTS component, we implement VITS 2, a single-stage end-to-end TTS system that generates high-quality, natural-sounding speech through adversarial learning and flow-based acoustic modeling. This work provides a scalable framework for developing Tibetan dialect conversion systems in under-resourced linguistic contexts.