Deep Feature Learning for Tibetan Speech Recognition using Sparse Auto-encoder

Abstract

HMM models based on MFCC features are widely used by researchers in Tibetan speech recognition. Although the shallow models of HMM are effective, they cannot reflect the speech perceptual mechanism in human being's brain. In this paper, we propose to apply sparse auto-encoder to learn deep features based on MFCC features for speech data. The deep features not only simulate sparse touches signal of the auditory nerve, and are significant to improve speech recognition accuracy with HMM models. Experimental results show that the deep features learned by sparse auto-encoder perform better on Tibetan speech recognition than MFCC features and the deep features learned by MLP.