Abstract:

F0 tracking is an essential part of the analysis of speech signals. There are many algorithms out there for determining the fundamental frequency of speech signals. Common issues are tracking errors, such as octave-jumps and other artefacts which do not accurately represent the speech signal. Some algorithms try to detect such unwanted discontinuities and perform automatic corrections. Sometimes, such corrections are unfortunately not of sufficient precision. For prosodic analysis, however, it is essential to avoid miscorrections of F0 that would lead to a distorted representation of one of the main acoustic correlates of interest in these investigations. Hence, we need to detect possible unwanted F0 tracking errors and make them processable for linguist experts instead of doing automatic corrections. The aim of this thesis is to implement a tool to detect, cluster and visualise possible unwanted discontinuities in F0 curves of speech signals.

The thesis can be found here (German).