Spectral tilt modelling with extrapolated GMMs for intelligibility enhancement of narrowband telephone speech

Estimating the spectral tilt of the glottal source from telephone speech using a deep neural network

“…Three techniques-stabilized weighted linear prediction (SWLP), 13 double linear prediction (2LP), 4 and cepstral fit (CF)-were used as reference methods that all estimate the spectral tilt of the glottal source directly from speech.…”

“…SWLP has been used previously for the mapping of the spectral tilt with Gaussian mixture models. 4 The parameter values used in this study were M ¼ 2 and p ¼ 6.…”

“…Spectral tilt has been analyzed, for instance, in studying linguistic stress, 1 phonation type, 2 vocal effort, 3 and speech intelligibility improvement. 4 While previous studies have measured the spectral tilt mainly in order to parameterize and classify speech production, tilt analysis can also be used to modify the speech signal, for instance, by changing its vocal effort characteristics. 4,5 For this type of application, a parametric representation, such as an all-pole model, that can be used to directly modify the tilt is desirable.…”

“…4 While previous studies have measured the spectral tilt mainly in order to parameterize and classify speech production, tilt analysis can also be used to modify the speech signal, for instance, by changing its vocal effort characteristics. 4,5 For this type of application, a parametric representation, such as an all-pole model, that can be used to directly modify the tilt is desirable. However, obtaining an all-pole model of the spectral tilt is not straightforward because direct, noninvasive acoustical recordings of the glottal source are not possible.…”

“…Therefore, previously all-pole techniques based on estimating the spectral tilt directly from the speech signal, instead of the true glottal excitation, have been used. 4 However, these methods might give tilt estimates that are distorted by poorly canceled formants.…”