Marcel Gabrea

A short introduction to wavelets and their applications

This simplified introduction to wavelets starts with a short historical background. The continuous wavelet transform is presented. The multiresolution concept is concisely described. It is followed by the filter banks method for wavelet analysis and reconstruction. Wavelet packets are briefly presented. Some typical applications are mentioned.

Robust adaptive Kalman filtering-based speech enhancement algorithm

The paper deals with the problem of speech enhancement when only a corrupted speech signal is available for processing. Kalman filtering is known as an effective speech enhancement technique, in which the speech signal is usually modeled as an autoregressive (AR) model and represented in the state-space domain. Various approaches based on the Kalman filter have been presented in the literature. They usually operate in two steps: first, additive noise and driving process statistics and speech model parameters are estimated and second, the speech signal is estimated by using Kalman filtering. In the paper, sequential estimators are used for suboptimal adaptive estimation of the unknown a priori driving process and of additive noise statistics simultaneously with the system state. The estimation of time-varying AR signal model is based on a robust recursive least-square algorithm with variable forgetting factor. The proposed algorithm provides improved state estimates at little computational expense.

Speech enhancement as a realisation issue

When enhancing a speech signal using a single microphone system, various approaches based on an autoregressive speech model are referenced in the literature. Using a Kalman filter, they operate in two steps: (1) the noise variances and the autoregressive parameters are estimated, (2) the speech signal is retrieved using standard Kalman filtering. However the existing methods are usually iterative and a voice activity detector (VAD) is often required to find the silent frames for the estimation of the variance of the white noise. To avoid these drawbacks, we propose to consider Kalman filter-based speech enhancement as a realisation issue, i.e. as the estimation of the system matrices in the state space representation using the estimation of the correlation function of the observations. For this purpose, we first present various solutions, based on works initially developed in the field of identification by Van Overschee et al. and Verhaegen. Their non-iterative extensions to coloured noise are also addressed and used with car noise. In the second part of the paper we propose an alternative approach based on Mehra et al. and Belangers approaches dealing with the estimation of the steady Kalman gain and previously derived in the framework of identification. This approach still avoids the use of a VAD.

Double affine projection algorithm-based speech enhancement algorithm

A symmetric feedback implementation scheme for two microphones speech enhancement is presented. We consider the coupling systems modeled as a linear time-invariant finite impulse response (FIR) filter and propose a new recursive-based adaptive filter solution to enhance the noisy speech. The optimum filter weight adaptation is based on a double affine projection algorithm (DAPA). This approach can be extended for a subclass of signal separations where the direct link is stronger than the interference link in both channels. A comparative study with other adaptive algorithms shows the superiority of the DAPA in performance improvement.

Wavelet based speech enhancement using two different threshold-based denoising algorithms

In this paper, we present a wavelet-based speech denoising technique obtained by the combination of the /spl mu/-law thresholding and the soft thresholding algorithm. Denoising is a compromise between the removal of the largest possible amount of noise and the preservation of signal integrity. To achieve a good implementation of this compromise we purpose the following procedure. The signal to be denoised is decomposed using wavelet packets up to the seventh level using DB11 wavelets. The /spl mu/-law thresholding is applied to all the final decomposition level subband coefficients except those of the two lower subbands on which soft thresholding is applied. To evaluate the performance of the proposed method, a clean speech dataset from the TIMIT database, corrupted with pink noise, for SNR levels ranging from 5 to 15 dB has been utilized. It has been found that the results obtained by our method are better than those given by each one of the two combined methods used separately.

Adaptive Kalman filtering-based speech enhancement algorithm

This paper deals with the problem of speech enhancement when only a corrupted speech signal is available for processing. Kalman filtering is known as an effective speech enhancement technique, in which the speech signal is usually modeled as an autoregressive (AR) model and represented in the state-space domain. Various approaches based on the Kalman filter are presented in the literature. They usually operate in two steps: first, additive noise and driving process variances and speech model parameters are estimated and second, the speech signal is estimated by using Kalman filtering. In this paper sequential estimators are used for suboptimal adaptive estimation of the unknown a priori driving process and additive noise statistics simultaneously with the system state. The estimation algorithm provides improved state estimates at little computational expense.

Automated pathologies detection in retina digital images based on complex continuous wavelet transform phase angles

An automated diagnosis system that uses complex continuous wavelet transform (CWT) to process retina digital images and support vector machines (SVMs) for classification purposes is presented. In particular, each retina image is transformed into two one-dimensional signals by concatenating image rows and columns separately. The mathematical norm of phase angles found in each one-dimensional signal at each level of CWT decomposition are relied on to characterise the texture of normal images against abnormal images affected by exudates, drusen and microaneurysms. The leave-one-out cross-validation method was adopted to conduct experiments and the results from the SVM show that the proposed approach gives better results than those obtained by other methods based on the correct classification rate, sensitivity and specificity.

An adaptive Kalman filter for the enhancement of speech signals in colored noise

This paper deals with the problem of speech enhancement when a corrupted speech signal with an additive colored noise is the only information available for processing. Kalman filtering is known as an effective speech enhancement technique, in which speech signal is usually modeled as autoregressive (AR) process and represented in the state-space domain. In the above context, all the Kalman filter-based approaches proposed in the past, operate in two steps: first, the noise and the signal parameters are estimated, and second, the speech signal is estimated by using Kalman filtering. In this paper a new sequential estimators are developed for sub-optimal adaptive estimation of the unknown a priori driving processes variances simultaneously with the system state. A weighted recursive least-square algorithm with variable forgetting factor is used for the estimation of the speech AR parameters and a recursively least-squares lattice algorithm is used for the estimation of the noise AR parameters. The algorithm provides improved speech estimate at little computational expense.

Correlation coefficient-based voice activity detector algorithm

A voice activity detector (VAD) is an algorithm able to distinguish the speech regions from the background noise of the input signal and is an important step in many speech processing applications. The varying nature and the large variety of speech and background noise make this problem difficult especially for low signal to noise ratio (SNR) that is the case for many practical applications. In this paper we propose a new VAD algorithm designed to improve the solution of word boundary detection problem for variable background noise level in a real time application. The input signal is windowed in time domain and then the energy and the spectrum of the current frame are obtained. The first few frames are supposed not to contain speech and are used to obtain a first estimate of the noise parameters. These parameters are updated during the silence periods using a first order autoregressive filter. In order to obtain robust parameters that do not depend on the amplitude of the spectrum, the correlation coefficient of the instantaneous spectrum and an average of the background noise spectrum is calculated. The speech regions may be detected based on a statistical approach using a simple binary Markov model for speech activity process. To evaluate the performance of the proposed method a clean speech dataset from the TIMIT database corrupted with different types of noise from NOISEX database for different SNR levels has been utilized.

Subspace state space model identification for speech enhancement

This paper deals with Kalman filter-based enhancement of a speech signal contaminated by a white noise, using a single microphone system. Such a problem can be stated as a realization issue in the framework of identification. For such a purpose we propose to identify the state space model by using subspace non-iterative algorithms based on orthogonal projections. Unlike estimate-maximize (EM)-based algorithms, this approach provides, in a single iteration from noisy observations, the matrices related to state space model and the covariance matrices that are necessary to perform Kalman filtering. In addition no voice activity detector is required unlike existing methods. Both methods proposed here are compared with classical approaches.