Pitch Range Estimation with Multi features and MTL-DNN Model
Qi Zhang, Chong Cao, Tiantian Li, Yanlu Xie, Jinsong Zhang · 2018
In speech communication, listeners can normalize the speaker's pitch by estimating the speaker's pitch range, even with a very short speech signal. Previous studies about pitch range estimation usually detected the maximum and minimum pitch values with large-scale speech samples through F0 distribution fitting. Based on the finding that spectral information played a role in the perception process of pitch range, this study used the multitask learning deep neural network (MTL-DNN) model to estimate pitch range from the perspective of spectral structure of a very short speech input. The results showed that, with the increase of the input speech length, the prediction performance was constantly improved. And the performance tended to be stable when the input speech length was about 300ms. The pitch range estimation performance of Fbank features outperformed that of MFCC features. With Fbank features, the mean absolute percentage errors (MAPE) and the mean absolute errors (MAE) of the ceiling was 10.16% and 24.57Hz respectively. And the MAPE and MAE of the floor was 11.39% and 11.77Hz respectively. When the pitch information was added to the Fbank and MFCC features, the pitch range prediction performance was not significantly improved if the acoustic input length was more than 300ms. The results showed the feasibility of pitch range estimation model and confirmed the importance of spectral information in pitch range estimation.