End-to-end Modeling for Chinese Dialect Speech Recognition

Chinese dialect recognition is a special case of general Mandarin recognition tasks, which is a challenging problem due to the less dialect data and time-consuming pronunciation model (PM). Meanwhile, the readability of the dialect recognition text is still far from satisfactory. In this paper, we propose an end-to-end model that directly transcribes the dialect speech to Mandarin text and demonstrate that the Mandarin recognition text can not only significantly improve the readability of recognition text, but also save time and expense without the need of the dialect PM. The experiments on the Taizhou (in Jiangsu province) dialect speech corpus show efficiency of the proposed Mandarin recognition system, and the feasibility of the Mandarin label for dialect recognition task.