A Robust Pitch Extractor Based on DTW Lines and CASA with Application in Noisy Speech Recognition

This paper proposes a robust pitch extractor with application in Automatic Speech Recognition and based on selecting pitch lines of a tonegram (a representation of the different pitch energies at each frame time). First, the tonegram and its maximum energy regions are extracted and a Dynamic Time Warping algorithm finds the most energetic trajectories or pitch lines from these regions. A second stage estimates the tonegram of the most energetic lines by applying Computational Auditory Scene Analysis rules which reject and group octave-related lines. The mean pitch of the speaker is estimated and the final pitch is estimated by rejecting lines which are outside from the mean pitch. The proposed pitch extractor is evaluated in a novel way - by means of the word accuracy of a Missing Data recognizer on Aurora-2 database.