Abdellah Kacha

Time-frequency analysis and instantaneous frequency estimation using two-sided linear prediction

This paper presents a new time-frequency distribution which uses a time-dependent two-sided linear predictor model. The current sample is estimated as a weighted sum of the past and future values. The two-sided linear prediction approach yields a smaller prediction error than that obtained by using the usual one-sided linear predictor model. To estimate the time-dependent coefficients of the two-sided linear predictor, these are expanded as a linear combination of a set of time functions basis which leads to an ensemble of equations of the type of Yule-Walker equations. The nonstationary power spectrum estimate is used as a time-frequency distribution to characterize the signal jointly in the time domain and the frequency domain. We show that two-sided prediction-based time-frequency distribution can discriminate two close components in the time-frequency plane that neither Choi-Williams distribution nor one-sided prediction-based time-frequency distribution are capable of resolving. Also, the proposed time-frequency distribution is used to estimate the instantaneous frequency. Examples show that the proposed approach outperforms the usual technique based on the nonstationary autoregressive model.

Multi-band dysperiodicity analyses of disordered connected speech

The objective is to analyse vocal dysperiodicities in connected speech produced by dysphonic speakers. The analysis involves a variogram-based method that enables tracking instantaneous vocal dysperiodicities. The dysperiodicity trace is summarized by means of the signal-to-dysperiodicity ratio, which has been shown to correlate strongly with the perceived degree of hoarseness of the speaker. Previously, this method has been evaluated on small corpora only. In this article, analyses have been carried out on two corpora comprising over 250 and 700 speakers. This has enabled carrying out multi-frequency band and multi-cue analyses without risking overfitting. The analysis results are compared to the cepstral peak prominence, which is a popular cue that indirectly summarizes vocal dysperiodicities frame-wise. A perceptual rating has been available for the first corpus whereas speakers in the second corpus have been categorized as normal or pathological only. For the first corpus, results show that the correlation with perceptual scores increases statistically significantly for multi-band analysis compared to conventional full-band analysis. Also, combining the cepstral peak prominence with the low-frequency band signal-to-dysperiodicity ratio statistically significantly increases their combined correlation with perceptual scores. The signal-to-dysperiodicity ratios of the two corpora have been separately submitted to principal component analysis. The results show that the first two principal components are interpretable in terms of the degree of dysphonia and the spectral slope, respectively. The clinical relevance of the principal components has been confirmed by linear discriminant analysis.

Estimation of dysperiodicities in disordered speech

This paper presents two methods for tracking vocal dysperiodicities in connected speech. The first is based on a long-term linear predictor with one coefficient and the second on a generalized variogram. Both analysis methods guarantee that a slight increase or decrease of irregularities in the speech signal produces a slight increase or decrease of the estimated vocal dysperiodicity trace. No spurious noise boosting occurs owing to erroneous insertions or omissions of speech cycles, or the comparison of speech cycles across phonetic boundaries. The two techniques differ with regard to how slow changes of speech cycle amplitudes are compensated for. They are compared on two speech corpora. One comprises stationary fragments of vowel [a] produced by 89 male and female normophonic and dysphonic speakers. Another comprises four French sentences as well as vowel [a] produced by 22 male and female normophonic and dysphonic speakers. Vocal dysperiodicities are summarized by means of global and segmental signal-to-dysperiodicity ratios. They are correlated with hoarseness scores obtained by means of perceptual ratings of the speech tokens. The two techniques obtain signal-to-dysperiodicity ratios that are statistically significantly correlated with the hoarseness scores. For connected speech, the segmental signal-to-dysperiodicity ratio correlates more strongly with perceptual scores of hoarseness than the global signal-to-dysperiodicity ratio.

Dysphonic speech analysis using generalized variogram

Acoustic analyses of speech signals are popular in the framework of the clinical evaluation of voice and the diagnosis of disease. We propose a new strategy for dysphonic speech analysis that extracts vocal dysperiocities by using a generalized form of the variogram. The generalized variogram allows the inherent drawbacks of both long-term and short-term linear prediction formulations, widely used in disordered speech analysis, to be overcome. The proposed approach uses a forgetting factor to account for the nonstationarity nature of the speech signal. Experimental results show that the proposed approach outperforms the double prediction-based technique.

Multiband frame-based acoustic cues of vocal dysperiodicities in disordered connected speech

Abstract Two variants of multiband segmental signal-to-dysperiodicity ratio are used to summarize vocal dysperiodicities in connected disordered speech and their performance is compared to that of the conventional global signal-to-dysperiodicity ratio. Acoustic analysis is carried out by means of a generalized variogram to extract vocal dysperiodicities. The corpus comprises four French sentences as well as vowels [a] produced by 22 male and female normophonic and dysphonic speakers. It is shown that the multiband signal-to-dysperiodicity ratios correlate better with perceptual scores of hoarseness than the global signal-to-dysperiodicity ratio. The highest correlations are achieved by the acoustic marker based on linear regression analysis of the segmental signal-to-dysperiodiciy ratios in different non-overlapping frequency bands. The perceptual scores are based on comparative judgments by six listeners of pairs of speech tokens.

Complex Blind Source Separation

Blind source separation (BSS) techniques aim at recovering the original source signals from observed mixtures without a priori information. The bivariate empirical mode decomposition (BEMD) algorithm combined with complex independent component analysis by entropy bound minimization (ICA-EBM) technique is proposed as an alternative to separate convolutive mixtures of speech signals. The empirical mode decomposition (EMD) is a local self-adaptive decomposition method that allows analyzing data from non-stationary and/or nonlinear processes. Its principle is based on the sequential extraction of different amplitude and frequency modulation single-component contributions called intrinsic mode functions (IMFs). The BEMD is an extension of the EMD to complex-valued signals. First, the convolutive mixtures in the frequency domain are decomposed into a set of IMFs using the BEMD algorithm, and then, the complex ICA-EBM method is applied to extract the independent sources. The performance of the proposed approach is tested on real speech sounds chosen from available databases and compared to the results obtained via conventional frequency ICA and BEMD-ICA-based separation for convolutive mixtures. Simulation results show that the proposed method of BSS outperforms the BEMD-ICA separation technique for convolutive mixtures and conventional frequency ICA.

Multiband vocal dysperiodicities analysis using empirical mode decomposition in the log-spectral domain

Abstract In this paper, empirical mode decomposition (EMD) is proposed as an alternative to decompose the log magnitude spectrum of the speech signal into its harmonic, envelope and noise components. The acoustic measure named harmonic-to-noise ratio (HNR) is used to summarize the degree of disturbance in the speech signal and consequently to evaluate the overall quality of the disordered voices produced by dysphonic speakers. Most approaches for HNR estimation have in common to involve the isolation of individual speech cycles or pseudo-harmonics/rhamonics in speech spectrum/cepstrum; however, this isolation cannot be carried out reliably in speech produced by severely hoarse speakers and may result in inaccurate HNR estimation. The EMD-based approach used in this study incorporates an appropriate procedure that estimates automatically the thresholds used by the clustering algorithm without knowledge of the fundamental frequency. The frequency range of the harmonic and noise components is divided into ten equally spaced intervals and the harmonic-to-noise ratios (HNRs) within each interval are used as independent variables to summarize the amount of perceived hoarseness. The proposed method is evaluated on a corpus comprising 251 normophonic and dysphonic speakers. Multiple correlation analysis carried out on HNRs from the different frequency bands shows that multi-band analysis based on empirical mode decomposition results in statistically significantly higher correlation of predicted scores with scores of perceived hoarseness over full-band analysis. Principal component analysis is carried out on the HNR measures obtained in the ten frequency bands. More than 97% of the total variance is explained by the first two principal components, PC1 and PC2. Experimental results show that the first principal component is interpretable in terms of the degree of the severity of hoarseness whereas the second principal component indicates whether the voice is high-pitched or low-pitched. It is shown that the first two principal components result in a high predictability of hoarseness scores.

Vocal dysperiodicities estimation by means of adaptive long-term prediction

An adaptive formulation of the long-term bi-directional linear predictive analysis is proposed in the context of the acoustic assessment of disordered speech. Vocal dysperiodicities are summarized by means of a signal-to-dysperiodicity ratio (SDR) marker. It is shown that performing an adaptive forward and backward long-term linear prediction of each speech sample and retaining the minimal prediction error energy as a cue of vocal dysperiodicity results in an SDR that correlates with the perceived degree of hoarseness. The coefficients of the time-varying long-term linear predictive model are estimated by means of the recursive least squares algorithm. The corpora comprise sustained vowels and French sentences produced by male and female normophonic and dysphonic speakers. A perceptual assessment of speech samples, which rests on comparative judgments, is used to evaluate the ability of the acoustic marker to predict subjective measures of voice quality. Experimental results show that the adaptive approach gives rise to high correlations for sustained vowels as well as for sentences.

Bivariate analysis of disordered connected speech using temporal and spectral acoustic cues

The presentation concerns the assessment of disordered voices produced by dysphonic speakers. The empirical mode decomposition algorithm is used to decompose the log of the magnitude spectrum of the speech signal into its harmonic, envelope and noise components and the harmonic-to-noise ratio (HNR) is used to summarize the overall quality of the disordered voices. The present study aims at improving a previously proposed algorithm by incorporating an appropriate method that estimates automatically the thresholds required by the algorithm without knowledge of the fundamental frequency and combining the temporal acoustic marker named segmental signal-to-dysperiodicity ratio (SDRSEG) with the harmonic-to-noise ratio in order to predict the degree of perceived hoarseness. The performances of the bivariate analysis-based approach for vocal dysperiodicities assessment in terms of correlation of the predicted perceived grade scores with the original perceived degree of hoarseness are investigated using a large corpus comprising concatenations of two Dutch sentences followed by vowel [a].

Analysis of Epileptic EEG Signals by Means of Empirical Mode Decomposition and Time-Varying Two-Sided Autoregressive modelling

The presentation concerns the analysis of EEG recordings for epileptic seizure detection. The EEG signal is decomposed adaptively into oscillating components called implicit mode functions (IMFs) using the empirical mode decomposition (EMD) algorithm and then a time-frequency analysis is carried out on the first two components using a parametric time-frequency distribution based on two-sided autoregressive modeling. The local frequency of the IMF extracted at some iteration is lower than that of the IMF extracted at the previous iteration which enables to analyze the signal at different scales. The relative variation of the instantaneous frequency of the IMFs estimated using two-sided autoregressive model-based time-frequency distribution is used as a feature for automatic seizure detection in the EEG recordings. The effectiveness of the proposed method is demonstrated on EEG signals recorded from 18 patients suffering from different kinds of epileptic seizures as well as on normal EEG data recorded from control population.