Investigations of low resource multi-accent mandarin speech recognition

The Mandarin speech always involves a rich set of regional accents, so that modeling the acoustic variabilities imposed by accents is a challenging task for Mandarin speech recognition. This work investigated using limited accented data to design a multi-accent decision tree, so as to improve the recognition accuracy of traditional GMM-HMM systems. Moreover, the deep neural networks with senone/monophone outputs were used with the multi-accent decision tree based GMM-HMM, in order to build up a robust tandem system. The experiments were evaluated on the database consisting of accented speech collected from seven typical accent regions. The systems designed with the proposed method significantly outperformed conventional GMM-HMMs systems by 1.66% absolute (8.1% relative). The tandem systems trained with DNN and multi-accent decision tree can further reduce the character error rate by 4.64% absolute (24.8% relative), compared to the accent-dependent GMM-HMM system.