Deep neural network acoustic modeling for native and non-native Mandarin speech recognition
Xin Chen, Jian Cheng · 2014
Recently, Context-Dependent Deep Neural Network Hidden Markov Models (CD-DNN-HMMs) proved to be a great success in improving the speech recognition accuracy over traditional Gaussian Mixture Model (GMM)-HMMs. In this paper, we applied CD-DNN-HMMs to an automatic Spoken Chinese Test (SCT) and improved the speech recognition performance significantly. We experimented on a large amount of native and non-native data collected for SCT and evaluated the effect of using different network activation functions: sigmoid and Rectified Linear Unit (ReLU) for acoustic modeling. We investigated accent adaptation on a Linear Input Network (LIN) - ReLU DNN network structure to enhance the non-native speech recognition accuracy. Furthermore, the feature-space Discriminative Linear Regression fDLR based accent adaptation was applied at the bottom of the system and was evaluated. In our task, we observed that applying prior probabilities was critical; pre-training did not help improve the performance; the performance using ReLU DNNs or sigmoid DNNs was similar. Overall, CD-DNN-HMM acoustic models made a relative 20% improvement on our non-native task, and a relative 47% improvement on the native task over traditional GMM-HMM models. The accent adaptation additionally contributed a 2.7% relative gain on the non-native task.