Time selected multiple algorithms for reliable f0 tracking in difficult recording conditions

Abstract

Prosodic data mining in large spontaneous speech corpora often requires acoustic analysis of recordings done in poor conditions. The most detrimental pertain to the absence first harmonic in voiced segments and the presence of echo. In such cases it is very difficult to separate inter vocalic consonantal voicing from echo, or recover Fo from remaining harmonics. To address these problems, 10 different pitch tracking algorithms were implemented in the software program WinPitch, in order to allow an expert user, guided by an underlying displayed narrow band spectrogram, to select manually the appropriate process to deliver a satisfactory Fo curve for a selected time segment.