Eric Grivel

Speech Enhancement Combining Optimal Smoothing and Errors-In-Variables Identification of Noisy AR Processes

In the framework of speech enhancement, several parametric approaches based on an a priori model for a speech signal have been proposed. When using an autoregressive (AR) model, three issues must be addressed. (1) How to deal with AR parameter estimation? Indeed, due to additive noise, the standard least squares criterion leads to biased estimates of AR parameters. (2) Can an estimation of the variance of the additive noise for each speech frame be obtained? A voice activity detector is often used for its estimation. (3) Which estimation rules and techniques (filtering, smoothing, etc.) can be considered to retrieve the speech signal? Our contribution in this paper is threefold. First, we propose to view the identification of the noisy AR process as an errors-in-variables problem. This blind method has the advantage of providing accurate estimations of both the AR parameters and the variance of the additive noise. Second, we propose an alternative algorithm to standard Kalman smoothing, based on a constrained minimum variance estimation procedure with a lower computational cost. Third, the combination of these two steps is investigated. It provides better results than some existing speech enhancement approaches in terms of signal-to-noise-ratio (SNR), segmental SNR, and informal subjective tests.

Dual

This paper deals with the joint signal and parameter estimation for linear state-space models. An efficient solution to this problem can be obtained by using a recursive instrumental variable technique based on two dual Kalman filters. In that case, the driving process and the observation noise in the state-space representation for each filter must be white with known variances. These conditions, however, are too strong to be always satisfied in real cases. To relax them, we propose a new approach based on two dual Hinfin filters. Once a new observation of the disturbed signal is available, the first Hinfin algorithm uses the latest estimated parameters to estimate the signal, while the second Hinfin algorithm uses the estimated signal to update the parameters. In addition, as the Hinfin filter behavior depends on the choice of various weights, we present a way to recursively tune them. This approach is then studied in the following cases: (1) consistent estimation of the AR parameters from noisy observations and (2) speech enhancement, where no a priori model of the additive noise is required for the proposed approach. In each case, a comparative study with existing methods is carried out to analyze the relevance of our solution.

{H}_{\infty}

The estimation of the parameters of an autoregressive process (AR) from noisy observations is still a challenging problem. In this paper, we propose to sequentially estimate both the signal and the parameters, avoiding a non-linear approach such as the extended Kalman filter. The method is based on two conditionally linked Kalman filters running in parallel. Once a new observation is available, the first filter uses the latest estimated AR parameters to estimate the signal, while the second filter uses the estimated signal to update the AR parameters. This approach can be viewed as a recursive instrumental variable-based method and hence has the advantage of providing consistent estimates of the parameters from noisy observations. A comparative study with existing algorithms illustrates the performances of the proposed method when the additive noise is either white or coloured.

Algorithms for Signal Processing— Application to Speech Enhancement

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.

Consistent estimation of autoregressive parameters from noisy observations based on two interacting Kalman filters

We propose a new estimation method for the characterization of the Hodgkin-Huxley formalism. This method is an alternative technique to the classical estimation methods associated with voltage clamp measurements. It uses voltage clamp type recordings, but is based on the differential evolution algorithm. The parameters of an ionic channel are estimated simultaneously, such that the usual approximations of classical methods are avoided and all the parameters of the model, including the time constant, can be correctly optimized. In a second step, this new estimation technique is applied to the automated tuning of neuromimetic analog integrated circuits designed by our research group. We present a tuning example of a fast spiking neuron, which reproduces the frequency-current characteristics of the reference data, as well as the membrane voltage behavior. The final goal of this tuning is to interconnect neuromimetic chips as neural networks, with specific cellular properties, for future theoretical studies in neuroscience.

Speech enhancement as a realisation issue

This paper deals with the identification of a nonlinear SISO system modelled by a second-order Volterra series expansion when both the input and the output are disturbed by additive white Gaussian noises. Two methods are proposed. Firstly, we present an unbiased on-line approach based on the LMS. It includes a bias correction scheme which requires the variance of the input additive noise. Secondly, we suggest solving the identification problem as an errors-in-variables issue, by means of the so-called Frisch scheme. Although its computational cost is high, this approach has the advantage of estimating the Volterra kernels and the variances of both the additive noises and the input signal, even if the signal-to-noise ratios at the input and the output are low.

Automated parameter estimation of the hodgkin-huxley model using the differential evolution algorithm: Application to neuromimetic analog integrated circuits

In 1952 Hodgkin and Huxley introduced the voltage-clamp technique to extract the parameters of the ionic channel model of a neuron. Although this method is widely used today, it has a lot of disadvantages. In this paper, we propose an alternative approach to the estimation method of the voltage-clamp technique using metaheuristics such as simulated annealing, genetic algorithms and differential evolution. This method avoids approximations of the original technique by simultaneously estimating all the parameters of a single ionic channel with a single fitness function. To compare the different methods, we apply them on measurements from a neuromimetic integrated circuit. This circuit, due to its analog behavior, provides us noisy data like a biological system. Therefore we can validate the efficiency of our method on experimental-like data.

Estimating second-order Volterra system parameters from noisy measurements based on an LMS variant or an errors-in-variables method

This letter deals with the identification of time-varying Rayleigh fading channels using a training sequence-based approach. When the fading channel is approximated by an autoregressive (AR) process, it can be estimated by means of Kalman filtering, for instance. However, this method requires the estimations of both the AR parameters and the noise variances in the state-space representation of the system. For this purpose, the existing noise compensated approaches could be considered, but they usually require a long observation window and do not necessarily provide reliable estimates when the signal-to-noise ratio is low. Therefore, we propose to view the channel identification as an errors-in-variables (EIV) issue. The method consists in searching the noise variances that enable specific noise compensated autocorrelation matrices of observations to be positive semidefinite. In addition, the AR parameters can be estimated from the null spaces of these matrices. Simulation results confirm the effectiveness of this approach, especially in presence of a high amount of noise.

Parameter estimation of the Hodgkin-Huxley model using metaheuristics: Application to neuromimetic analog integrated circuits

Autoregressive (AR) and time-varying AR (TVAR) models are widely used in various applications, from speech processing to biomedical signal analysis. Various dissimilarity measures such as the Itakura divergence have been proposed to compare two AR models. However, they do not take into account the variances of the driving processes and only apply to stationary processes. More generally, the comparison between Gaussian processes is based on the Kullback-Leibler (KL) divergence but only asymptotic expressions are classically used. In this letter, we suggest analyzing the similarities of two TVAR models, sample after sample, by recursively computing the Jeffreys divergence between the joint distributions of the successive values of each TVAR model. Then, we show that, under some assumptions, this divergence tends to the Itakura divergence in the stationary case.

Errors-In-Variables-Based Approach for the Identification of AR Time-Varying Fading Channels

In various applications from radar processing to mobile communication systems based on CDMA or OFDM, M-AR multichannel processes are often considered and may be combined with Kalman filtering. However, the estimations of the M-AR parameter matrices and the autocorrelation matrices of the additive noise and the driving process from noisy observations are key problems to be addressed. In this paper, we suggest solving them as an errors-in-variables issue. In that case, the noisy-observation autocorrelation matrix compensated by a specific diagonal block matrix and whose kernel is defined by the M-AR parameter matrices must be positive semi-definite. Hence, the parameter estimation consists in searching every diagonal block matrix that satisfies this property, in reiterating this search for a higher model order and then in extracting the solution that belongs to both sets. A comparative study is then carried out with existing methods including those based on the Extended Kalman Filter (EKF) and the Sigma-Point Kalman Filters (SPKF). It illustrates the relevance and advantages of the proposed approaches.