The potential of acoustic measures for measuring the supranormal speaking voice

Abstract

INTRODUCTION: Acoustic analysis is commonly used to objectively measure the voice as it is non-invasive and in-expensive. Many acoustic measures have been shown to reliably measure quasi-periodic voice signals. Normal and supranormal (better-than-normal) voices may therefore be reliably analyzed with a range of acoustic measures however it is not known which of these measures may distinguish supranormal from normal voices. METHOD: Sixty-nine female participants (mean age = 20.44 years, range = 17-25) were recorded while producing a sustained vowel /a/. All samples were classified as Type 1/quasi-periodic (Titze, 1995). The samples were then analyzed for six acoustic noise measures: Noise-to-Harmonic Ratio (NHRP) and Harmonic-to-Noise Ratio (HNR) in Praat; Glottal-to-Noise Ratio (GNE) in LingWaves; Noise-to-Harmonic Ratio (NHRM) in the Multi-Dimensional Voice Program; and Cepstral Peak Prominence (CPP) and smoothed Cepstral Peak Prominence (CPPS) in SpeechTool. Results were converted to z-scores and frequency histograms were calculated for each acoustic measure. Distribution for each measure was calculated using skewness and kurtosis. RESULTS: The distribution of the histograms for GNE (skewness = -0.955, S.E. = 0.289, p = 0.001) and NHRP (skewness = -1.041, S.E. = 0.289, p = 0.0004) were significantly positively skewed. Distribution for GNE was also kurtotic (kurtosis =1.36, S.E. = 0.57, p = 0.017). The histograms for all other variables were normally distributed. DISCUSSION: Measures of CCP, CPPS, NHRM and HNRP have potential to measure and distinguish supranormal voices. This is because participants’ results spread normally across the measures’ full ranges, allowing sufficient discrimination to distinguish supranormal voices. Due to the positive skewness seen for GNE and NHRP, they may not be useful for distinguishing supranormal from normal voices. GNE produced a ceiling effect when measuring better-than-normal voices, as many voices were analyzed close to the maximum of +1.