Multitaper harmonic analysis of infant vocalizations

Abstract

Infant vocalizations exhibit many special complexities relative to adult speech that make acoustic analysis difficult, even though the basic physics of sound production in the vocal tract is identical. Fundamental frequency is raised, resulting in wider-spaced harmonics that make it hard to accurately estimate the spectral envelope or locate formant resonances, which are higher than in adult voices, due to anatomical scaling effects. Irregular vocal fold vibratory regimes are often observed, involving period doubling and other nonlinear phenomena not seen in adult speech, along with atypical patterns of turbulent noise production. Unsurprisingly, traditional analysis techniques that depend on estimating the spectral envelope from the speech signal do not work well when applied to infant vocalizations. An alternative approach is proposed, using multitaper analysis to calculate the time-varying amplitude and phase of every harmonic component of the voice, along with the residual noise component, recovering the vocal tract transfer function from the results. The new technique is compared against linear prediction and cepstral analysis using home audio recordings of 20 infants collected from 0 to 24 months using LENA technology. Developmental progressions in the acoustic structure of the infant voice are identified that cannot be found using traditional methods.