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

Abstract

Estimation of the spectral tilt of the glottal source has several applications in speech analysis and modification. However, direct estimation of the tilt from telephone speech is challenging due to vocal tract resonances and distortion caused by speech compression. In this study, a deep neural network is used for the tilt estimation from telephone speech by training the network with tilt estimates computed by glottal inverse filtering. An objective evaluation shows that the proposed technique gives more accurate estimates for the spectral tilt than previously used techniques that estimate the tilt directly from telephone speech without glottal inverse filtering.

Evaluation of Spectral Tilt Measures for Sentence Prominence Under Different Noise Conditions

“…For each signal, the standard acoustic correlates for prominence, namely, energy, F0, and duration were computed (section 2.1) together with several tilt measures that have been commonly used in the literature (section 2.2). In addition, a DNN-based spectral tilt estimation was also evaluated (similar to [26]) in order to investigate (i) the efficiency of DNN-based source tilt estimation for prominence, and (ii) the potential for the DNN to add robustness on tilt estimation for noisy signals (section 2.3). For all features, a number of aggregate statistical measures over words were then computed, and their capability to discriminate prominent from non-prominent words was measured in terms of the separability of the feature distributions.…”

“…A new method was proposed recently in [26] to estimate and parameterize the glottal source spectrum in noisy, non-ideal conditions where conventional GIF analysis cannot be used due to its known sensitivity to noise [28]. The method proposed in [26] uses a deep neural network (DNN) to map an input feature vector (the logarithmic speech power spectrum) into an output vector (all-pole model of the glottal source spectrum parameterized using line spectrum frequencies (LSFs)).…”

“…The method proposed in [26] uses a deep neural network (DNN) to map an input feature vector (the logarithmic speech power spectrum) into an output vector (all-pole model of the glottal source spectrum parameterized using line spectrum frequencies (LSFs)). In this work, two separate standard feed-forward DNNs were trained for the prediction of the LSFs describing the glottal source spectrum directly from the logarithmic FFT magnitude spectrum of the speech input (20-ms window, 10-ms hop size; see also Fig.…”

“…As the methods for the estimation of spectral tilt are highly variable, the goal of the present work is to compare a range of the most well-known measures for spectral tilt together with the newly-proposed deep neural net (DNN) -based technique [26]. The capability of the spectral tilt measures to describe prosodic prominence is then evaluated using clean and noisecorrupted Dutch speech.…”

“…In addition, some studies utilize measures such as regression line fitting and harmonic ratio, but, instead of applying the measures directly on the short-term spectrum of speech, they utilize the spectrum of the glottal source waveform that is obtained through glottal inverse filtering (GIF) [24,25]. Recently, a DNN-based system for robust spectral tilt estimation was described in [26], enabling the estimation of the glottal source tilt in non-ideal signal conditions without explicitly using GIF in the estimation phase.…”