Chinese Dialect Recognition Using X-vector with Multi-task Learning

Abstract

Speech recognition is often hard for language in countries with diverse accents, like Mandarin. However, the recognition of different dialects may help to improve the later speech recognition. In this paper, we conduct some experiments on a ten-dialect recognition task. We improve the system with different methods, like x-vector and multi-task learning with phone recognition. Finally, with the combination of the methods, we improve the accuracy of baseline model from 76.48% to 88.25%, where the relative improvement is 15%.