KNOWLEDGE TRANSFER FOR RUSSIAN CONVERSATIONAL TELEPHONE AUTOMATIC SPEECH RECOGNITION

Abstract

This paper describes the method of knowledge transfer between the ensemble of neural network acoustic models and student-network. This method is used to reduce computational costs and improve the quality of the speech recognition system. The experiments consider two variants of generation of class labels from the ensemble of models: interpolation with alignment, and the posteriori probabilities. Also, the quality of models was studied in relation with the smoothing coefficient. This coefficient was built into the output log-linear classifier of the neural network (softmax layer) and was used both in the ensemble and in the student-network. Additionally, the initial and final learning rates were analyzed. We were successful in relationship establishing between the usage of the smoothing coefficient for generation of the posteriori probabilities and the parameters of the learning rate. Finally, the application of the knowledge transfer for the automatic recognition of Russian conversational telephone speech gave the possibility to reduce the WER (Word Error Rate) by 2.49%, in comparison with the model trained on alignment from the ensemble of neural networks.

Keywords

• knowledge transfer
• smoothing coefficient
• softmax
• automatic speech recognition
• ensemble of neural networks
• student-network
• conversational telephone speech