Noise Robust Fundamental Frequency Estimation of Speech using CNN-based discriminative modeling

The fundamental frequency (F0) is a quantity representing the pitch of periodic signal and its estimation for time-variant quasiperiodic acoustic signal is one of common problems in speech processing studies. The correct estimation of this contributes to the improvement of speech processing systems such as, analysis of prosody, test-to-speech system and speech recognition system. While many algorithms have been proposed and they exhibit excellent performance for clean environment, it is a very difficult task for noisy environment. It is generally known that machine learning approach is effective as a discriminative model for handling data in which noise is mixed. In this paper, we propose a robust fundamental frequency estimation method for noisy speech signal by using convolutional neural network (CNN) which is a type of deep neural network (DNN). In our proposed method, convolution layer and pooling layer serve as an approximator of autocorrelation analysis and followed by discriminative modeling for classifying quantized F0 state. This process acquires a discriminator that extracts noise robust FO features. Experimental result showed that our method outperforms convolutional methods based on autocorrelation analysis and its combination with DNN modeling.