Automatic classification of laryngeal mechanisms in singing based on the audio signal

Abstract

Laryngeal mechanisms are physical adjustments of the vocal tract that enable the human voice capacity of producing a wide frequency range. They are also directly related to the notion of registers, a common concept in singing that may be defined as different vocal qualities or timbre changes by the singers. Thus, laryngeal mechanisms are an important parameter of human voice production, both for speech and for singing. Until now, the identification of laryngeal mechanisms is a manual labor. A trained clinician needs to analyze the outcomes of electroglottography (a specific clinical procedure) to make the distinction among these mechanisms. This paper presents a method for the automatic classification of laryngeal mechanisms. The approach consists of calculating textural features extracted from the images of the spectrogram of the audio signal and making the distinction among the mechanisms using a classifier. Our proposal achieved promising results without parameter optimization. This fact indicates that it is suitable to perform the discrimination between laryngeal mechanisms employing the visual aspect of spectrograms. Besides, there are no other work in literature concerning about automatically classifying these mechanisms, and further, based only on the voice signal.