FiPPiE: A Computationally Efficient Differentiable method for Estimating Fundamental Frequency From Spectrograms

Lev Finkelstein, Chun-an Chan, Vincent Wan, Heiga Zen, Rob Clark

Keywords: Speech, Text-to-Speech, Pitch Tracking, Frequency Estimation, Signal Processing

TL;DR: We present FiPPiE, a Filter-Inferred Pitch Posteriorgram Estimator -- a method of estimating fundamental frequency from spectrograms, either linear or mel, by applying a special kind of filter in the spectral domain.

Abstract: In this paper we present FiPPiE, a Filter-Inferred Pitch Posteriorgram Estimator – a method of estimating fundamental frequency from spectrograms, either linear or mel, by applying a special kind of filter in the spectral domain. Unlike other works in this field, we developed a procedure for training an optimized filter (or kernel) for this type of estimation. FiPPiE, based on this optimized filter, demonstrated itself as a reliable fundamental frequency estimator that is computationally efficient, differentiable, and easily implementable. We demonstrate the performance of the method both by the analysis of its behavior on human recordings, and by the stability analysis with help of an automated system.