Automatic estimation of the first three subglottal resonances from adults’ speech signals with application to speaker height estimation

Abstract

Recent research has demonstrated the usefulness of subglottal resonances (SGRs) in speaker normalization. However, existing algorithms for estimating SGRs from speech signals have limited applicability—they are effective with isolated vowels only. This paper proposes a novel algorithm for estimating the first three SGRs (Sg1,Sg2 and Sg3) from continuous adults’ speech. While Sg1 and Sg2 are estimated based on the phonological distinction they provide between vowel categories, Sg3 is estimated based on its correlation with Sg2. The RMS estimation errors (approximately 30, 60 and 100Hz for Sg1,Sg2 and Sg3, respectively) are not only comparable to the standard deviations in the measurements, but also are independent of vowel content and language (English and Spanish). Since SGRs correlate with speaker height while remaining roughly constant for a given speaker (unlike vocal tract parameters), the proposed algorithm is applied to the task of height estimation using speech signals. The proposed height estimation method matches state-of-the-art algorithms in performance (mean absolute error=5.3cm), but uses much less training data and a much smaller feature set. Our results, with additional analysis of physiological data, suggest the existence of a limit to the accuracy of speech-based height estimation.