Identifying highly rhythmic stretches of speech with envelope-driven resonance analysis

Abstract

This paper proposes envelope-driven resonance analysis as a technique for characterizing rhythmicity in speech, emphasizing the degree to which a brief stretch of speech creates a rhythmic expectancy. Most approaches to characterizing the rhythm of speech have utilized measurements derived from the durations of linguistically relevant units such as feet, syllables, or vocalic/consonantal intervals. Recently, alternative approaches have been developed which are based upon the amplitude envelope of the speech waveform after the waveform has been filtered to emphasize low-frequency oscillations associated with alternations between vowels and consonants. These approaches include spectral analysis and empirical mode decomposition of the envelope. The method explored here is resonance analysis, which utilizes a bank of resonators that differ in their characteristic resonant frequencies. The resonators are 2nd order dynamical systems analogous to driven, damped springs. The powers of the resonator amplitudes are analyzed during and subsequent to excitation by a speech amplitude envelope. The power and frequency distribution of the resonant response is used to identify highly rhythmic stretches of speech and characterize their spectral properties.