Weighted Autocorrelation with Convolutional Neural Network for Noisy Speech Pitch Estimation
Theint Theint Nway, Tetsuya Shimamura · 2024
In this paper, we propose a pitch estimation method for speech in noisy environments, which is an extended version of the state-of-the art methods based on deep learning; CREPE and hf0. Unlike the conventional methods, in the proposed method, many layers and parameters of convolutional neural network (CNN) are not required. Avoiding a repeated calculation of autocorrelation function, only two layers of CNN are utilized. A transformation of the input speech signal to its weighted autocorrelation function is simply and effectively employed for the input to the CNN. Experimental results show that the proposed method can achieve more accurate pitch estimation than the CREPE and hf0 do in noisy environments.