关键词: 越南语, 文本正则化, 编辑对齐算法, 语音合成

Abstract: Text normalization is a crucial pre-processing step in text-to-speech synthesis front-end analysis. The Viet-namese language presents challenges related to the semantic ambiguity of non-standard words such as num-bers and dates. Neural text normalization systems can leverage context; however, they suffer from unrecov-erable errors. This study introduces an end-to-end text normalization method grounded in edit constraints. Taking full account of the linguistic characteristics of Vietnamese, we propose a specialized text normalization annotation method for Vietnamese, aiming to enhance the model's contextual semantic information modeling. Additionally, an edit alignment algorithm is applied to effectively restrict the scope of non-standard word text, thereby reducing the search space during decoding. This mitigates text normalization prediction errors arising from inherent model limitations. The experimental results show 97% accuracy in Vietnamese text normali-zation using the proposed method. Moreover, its effectiveness extends to an open-source Chinese dataset, validating the applicability of the method beyond the Vietnamese language.

Key words: Keywords: Vietnamese, Text normalization, Edit alignment algorithm, TTS