Objective measures of breathy voice quality obtained using an auditory model

Abstract

While several acoustic measures have been proposed to quantify listener ratings of breathy voice quality, most have failed to give a consistent and high correlation with perceptual ratings of breathiness. One reason for these limitations is that most acoustic measures do not address the nonlinear processes that occur in the peripheral auditory system during the auditory perceptual process. It was hypothesized that modeling such nonlinear events during signal processing may provide objective parameters that better correspond to perceptual ratings of breathy voice quality. Ten listeners rated 27 voice stimuli using a five-point rating scale. Acoustic measures were determined from these stimuli and were selected based on their history of having a moderate to strong correlation to perceptual ratings of breathiness. The stimuli were also analyzed using an auditory model proposed by Moore, Glasberg, and Baer [J. Audio Eng. Soc. 45(4), 224-239 (1997)], and new measures were calculated from the output of this model. These measures included the partial loudness of the signal and the loudness of the aspiration noise. Measures obtained from the output of the auditory model were found to account for a high amount of variance in the perceptual ratings of breathiness.

Perceptual interaction of the harmonic source and noise in voice

“…Currently available measures of the NHR and/or of the slope of the harmonic voice source do not reflect context effects on perceptual acuity. As a result, the same NHR value measured from two voices with different harmonic sources may correspond to very different levels of perceived "noisiness," while voices with rather different These results differ somewhat from previous studies using an "auditory model" to quantify the acoustic precursors of breathy voice quality (Shrivastav and Sapienza, 2003). That model accounts for variations in perceived breathiness via two measures: the partial loudness of the harmonic source (as masked by noise) and the loudness of the noise source, unmasked by harmonic energy (Moore et al, 1997).…”

“…This further limits the extent to which results of such studies can be meaningfully generalized to more complex speech stimuli. In fact, in one study, application of the auditory model increased predictive power by only 4% as compared to traditional acoustic measures (Shrivastav and Sapienza, 2003), possibly as a result of the factors discussed above.…”

A model for the prediction of breathiness in vowels

“…Since this approach was not feasible when using natural stimuli, the AH was estimated using an algorithm described by Milenkovic (1995Milenkovic ( , 1997 and implemented in the software CSPEECH (Milenkovic; University of Wisconsin-Madison, WI). This algorithm has been used in prior experiments on breathiness (e.g., Shrivastav and Sapienza, 2003). Briefly, this algorithm attempts to determine a perfectly periodic glottal source within a small temporal window and estimates the AH by subtracting the actual vowel waveform from the ideal waveform.…”

“…These two measures computed from the output of this loudness model were observed to correlate with perceptual judgments of breathiness. Breathiness was observed to be inversely related to PL and related directly to NL (Shrivastav, 2003;Shrivastav and Sapienza, 2003). Shrivastav and Sapienza (2003) reported that for low to moderate perceptual judgments of breathiness, PL had the most predictive leverage, but NL better predicted breathiness for stimuli judged to have high levels of breathiness.…”

“…More recently, an auditory-processing front-end has been used as a preprocessing step prior to computing acoustic correlates for breathiness (Shrivastav, 2003;Shrivastav and Sapienza, 2003). This intermediate step represents the non-linear transformation between the acoustic signal and perception following processing by the auditory system.…”