Speech rhythm analysis with empirical mode decomposition
Sam Tilsen · The Journal of the Acoustical Society of America · 2011
This paper presents a new approach to characterizing speech rhythm, based upon empirical mode decomposition (EMD) [Huang etal., Proc. R. Soc. London Ser. A 454, 903–995 (1998)]. Intrinsic mode functions (IMFs) are obtained from EMD of a vocalic energy envelope of speech, which is a smoothly varying signal, representing primary vocalic resonance energy. EMD uses an iterative sifting process to decompose the signal into IMFs with zero mean and zero-crossings between extrema. Investigations of IMFs revealed that the last two IMFs appear to capture foot- and syllable-timescale oscillations in the envelope, respectively. EMD was applied to a seven-language corpus of speech previously used to compare languages using interval-based rhythm metrics [Ramus etal., Cognition 73, 265–292 (1999)]. The ratio of signal power in the foot- and syllable-associated IMFs was used as metric of the relative influence of foot-based timing on speech. The performance of the IMF power ratio metric in distinguishing languages is comparable to analyses based on interval metrics. The method is argued to have the potential for broad applicability in speech research.