A general framework for multi-accent Mandarin speech recognition using adaptive neural networks

In this paper, we propose a general framework for multi-accent speech recognition that combines Multi-level Adaptive Network (MLAN) and automatic model selection system based on accent classification. This framework solves the problem of domain mismatch between standard Mandarin and accent data and makes full use of limited accent data. The effectiveness of the proposed method was evaluated on two typical Chinese accent data, namely Shanghai and Chongqing accents. Results show higher performance of the framework on multi-accent speech recognition compared to GMM-HMM systems with prior accent label knowledge, with up to 3.89% CER (Character Error Rate) reduction on Chongqing accent test set and 1.71% on Shanghai accent test set.