Analysis of autocorrelation-based parameters in Creaky voice

Abstract

“Creaky voice” has many other terminologies, such as “creak”, “vocal fry”, “glottal fry”, “laryngealization”, “glottalization”, and “pulse register phonation”, used in several research areas (like Linguistics, Physiology, and Phonetics) [1,2]. Creaky voice is defined as “... a train of discrete laryngeal excitations, or “pulses”, of extremely low frequency (7 to about 78 Hz), with almost complete damping of the vocal tract between excitations.” (Hollien 66 cited in [1]). “The auditory effect is of a rapid series of taps, like a stick being run along a railing.” (Catford 64 cited in [2]). Creaky phonation carries many linguistic and paralinguistic information, depending on the language. For example, contrast between creaky and modal voicing among vowels and nasals is particularly common in some American Indian languages [3]. In [4,5], relationship between different phonation types and paralinguistic information like emotions and attitudes are reported for English. Strong correlations were reported between creaky voice and perception of relaxed/stressed, sad/happy, and bored/interested. In Japanese, expressive pressed voice that is frequently realized by creaky phonation also carries important paralinguistic information such as attitudes, emotional states and emphasis [6]. Further, in creaky segments, periodicity is disturbed and the pitch extraction becomes difficult, affecting the subsequent prosodic analysis, like intonation. Tendency of creaky segments for specific tone types is reported in [7] for phrase finals in Japanese. In the JST/CREST ESP Project [8], one of the goals is an expressive speech synthesizer based on unit selection, using a large database of spontaneous speech. For this purpose, labels of voice qualities (phonation types) become as important as prosodic labels. With the goal of doing automatic labeling of voice quality on a large speech database, in the present research, we focus on the automatic detection of creaky phonation.