Pitch detection in pathological voices driven by three tailored classical pitch detection algorithms

Pitch detection is one of the most difficult problems encountered when analyzing speech signals. This paper focuses on detecting the pitch in pathological voices, what is of key importance, for voice pathology diagnosis. In particular, we put special emphasis on diagnosing three common pathologies: roughness, hoarseness and breathiness. The intelligibility and quality of pathological voices are usually degraded by insufficient vocal power and abnormally low and unstable pitch, what basically involves detecting the pitch in this type of signals is a challenging goal. In this respect, in this paper we propose the use of three classical robust algorithms tailored for driving a particular sort of pitch detection in pathological voices. These methods are as follows: 1) an algorithm based on autocorrelation, 2) the so-called Harmonic Product Spectrum, and finally 3) an approach based onWavelet Transform. We have chosen these methods because they are well-known in pitch detection in healthy voices in different domains. In the effort of clearly showing to what extent the use of these three methods are appropriate for pitch detection in pathological voices, we have compared the results obtained with these three methods with those achieved by a commercial software for the same pathological voices database.