Analyses of voice and glottographic signals in singing and speech

Abstract: Recent advances in machine learning and time series analysis techniques have brought new perspectives to a great number of scientific fields. This thesis contributes applications of such techniques to voice analysis, in an attempt to extract information on the vibration of the vocal folds as such, as well as on the radiated acoustic signal. The data that was analyzed in this work are acoustic recordings, electroglottographic (EGG) signals and transnasal high- speed videoendoscopic images. The data analysis techniques are primarily based on clustering, i.e., grouping of data based on similarity, and sample entropy analysis, i.e., quantifying the degree of irregularity in a given signal. The experiments were conducted so as to provide data for different types of vibratory behaviors (or vibratory states) of the vocal folds. Clustering was used in order to categorize in an unsupervised fashion these different vi- bratory states, based solely on the electroglottographic signal, or the glottal area waveform, or both. Sample entropy was utilized as an indicator of in- stabilities, when subjects produced voiced sounds using irregular vibratory patterns, such as register breaks, intermittent diplophonia, and other types of irregularities. The prominent role of sound pressure level and fundamental frequency motivated further study of the relationship between them and the shape of the electroglottographic waveform. Graphical representations were created to visualize the relationship between different vibratory behaviors with fundamental frequency and sound pressure level. The EGG waveform shape was seen to depend strongly on sound pressure level and somewhat less on fundamental frequency. In very soft phonation, the almost sinusoidal waveform of the EGG suggests that studying the EGG using clusters may give a better representation compared to conventional time-domain metrics. The paradigm of the clustering was later applied in synchronous recordings of electroglottogram and glottal area waveforms in professional tenor singers. Different vibratory states were classified successfully using clustering, and the electroglottogram was seen to be as good as the glottal area waveform for such a classification task. The last part of this work concerns voices from subjects with organic dysphonia. A study was dedicated to investigate how vowel context (sustained versus excerpted from speech) can affect the power of quantitative acoustic measures to discriminate dysphonic subjects from controls. Two acoustic voice quality measures were used: the cepstral peak prominence (smoothed) and sample entropy. The cepstral peak prominence (smoothed) showed better discriminatory power with excerpted vowels, while sample entropy with sustained vowels. Additionally, it was found that sample entropy was strongly correlated with cepstral peak prominence (smoothed) and with the perceptual quality of breathiness.