Characterizing glottal activity from speech using Empirical Mode Decomposition

Glottal activity is an important aspect of speech production that results in voiced speech, and localizing such regions for computing various parameters of the excitation source is useful in many speech processing applications. The aim of this paper is to investigate the ability of Empirical Mode Decomposition (EMD) and its noise assisted variants, in characterizing glottal activity from the speech signal. A pair of consecutive Intrinsic Mode Functions (IMFs), obtained from the decomposition is found to reflect the periodic nature of different voiced regions of the speech signal. This IMF pair is utilized to construct a signal, named the Glottal Intrinsic Mode Function (GIMF), which represents most of the voiced speech regions. To measure the capability of the GIMF in representing the glottal activity, it is applied to the tasks of Glottal Activity Detection (GAD), pitch frequency (F₀) tracking and detecting pitch markers. The results ascertain the capability of EMD in localizing Glottal activity within a small subset of IMFs, and suggest the possibility of accurately extracting source-information from voiced speech with simple signal processing procedures.