A method of long-short time Fourier transform for estimation of fundamental frequency

Abstract:
The estimation of F0, or Fundamental Frequency, is one of the most vital steps of preprocessing research in speech and signal processing. Nevertheless, for small sample size problems, traditional methods and methods based on machine learning have limitations. This paper proposed a method for F0 estimation which integrates FFT, Band-pass Filtering (BPF) and STFT. In this method, the spectrum peaks are used in one hand to set frequency band of Band-pass Filtering, in another hand to calculate the proper length of frames which is used for STFT analysis. This greatly improved the performance of F0 estimation. The results show that, compared with Auto-Correlation and STFT method, the RMSE of the proposed LSTFT method has been reduced by 62% and 78%, respectively. This study indicates that the accuracy of F0 estimation used LSTFT method performs better than that of traditional method, LSTFT makes full use of the spectrum of the whole signal, and makes the short-time analysis such as STFT more dynamic.