Robust Multi-Pitch Tracking: a trained classifier based approach

Pitch determination algorithm (PDA) performance typically degrades in the presence of interfering speakers and other periodic sources. We propose a multi-pitch determination algorithm (MPDA) that will detect and estimate two pitch tracks, a dominant and an interferer. Our method aims at being robust to various levels of combination of two speakers. Similar to the Subband Autocorrelation Classification (SAcC) method, we present a classifier based approach trained on compressed correlogram features. In contrast we train our classifier to detect all periodic sources, allowing for multiple speakers to be present. Viterbi decoding over a Markov chain of possible pitch and multiple speaker states is used to generate significant and continuous pitch tracks. We will compare our proposed algorithm against another MPDA and evaluate the performance of the methods with metrics extended from traditional PDAs.
 

Speech