Estimação de Frequência Fundamental de Sinais Acústicos Ruidosos com Aprendizado de Máquina

This letter presents a proposal in the time domain to improve the estimates of the fundamental frequency (F 0) of the HHT-Amp method in noisy speech signals. The voiced frames are classified as high/low frequency by a Deep Convolutional Neural Network (DCNN), and candidates are extracted according to the most probable types of estimation errors. Finally, a cost function is defined as criterion for selecting the new value of F 0. The results of the experiments showed the superiority of the proposed solution DCNN+HHT-Amp in different scenarios when compared to competitive methods.