Linear Prediction Analysis Of Speech Research Articles

Quasi-closed phase forward-backward linear prediction analysis of speech for accurate formant detection and estimation

Recently, a quasi-closed phase (QCP) analysis of speech signals for accurate glottal inverse filtering was proposed. However, the QCP analysis which belongs to the family of temporally weighted linear prediction (WLP) methods uses the conventional forward type of sample prediction. This may not be the best choice especially in computing WLP models with a hard-limiting weighting function. A sample selective minimization of the prediction error in WLP reduces the effective number of samples available within a given window frame. To counter this problem, a modified quasi-closed phase forward-backward (QCP-FB) analysis is proposed, wherein each sample is predicted based on its past as well as future samples thereby utilizing the available number of samples more effectively. Formant detection and estimation experiments on synthetic vowels generated using a physical modeling approach as well as natural speech utterances show that the proposed QCP-FB method yields statistically significant improvements over the conventional linear prediction and QCP methods.

An efficient solution to sparse linear prediction analysis of speech

We propose an efficient solution to the problem of sparse linear prediction analysis of the speech signal. Our method is based on minimization of a weighted l2-norm of the prediction error. The weighting function is constructed such that less emphasis is given to the error around the points where we expect the largest prediction errors to occur (the glottal closure instants) and hence the resulting cost function approaches the ideal l0-norm cost function for sparse residual recovery. We show that the efficient minimization of this objective function (by solving normal equations of linear least squares problem) provides enhanced sparsity level of residuals compared to the l1-norm minimization approach which uses the computationally demanding convex optimization methods. Indeed, the computational complexity of the proposed method is roughly the same as the classic minimum variance linear prediction analysis approach. Moreover, to show a potential application of such sparse representation, we use the resulting linear prediction coefficients inside a multi-pulse synthesizer and show that the corresponding multi-pulse estimate of the excitation source results in slightly better synthesis quality when compared to the classical technique which uses the traditional non-sparse minimum variance synthesizer.

A GEOSTATISTICAL MODEL FOR LINEAR PREDICTION ANALYSIS OF SPEECH

This paper presents a geostatistical model as a new approach to the linear prediction analysis of speech. The autocorrelation method of autoregressive modeling, which is widely applied in the linear predictive coding of speech, is used as a benchmark for comparison with the present algorithm. Before discussing the proposed model, we will briefly describe the concepts of linear prediction analysis of speech and how this is solved by the well-known method of autocorrelation. Following is the introduction of geostatistics including the ideas of regionalized variables, semi-variograms and kriging equations. We then propose a geostatistical model to the linear prediction modeling of speech signals. Examples on speech data are given to illustrate the effectiveness of the present algorithm in comparison with the autocorrelation method. Advantages offered by the proposed geostatistical algorithm over the autocorrelation method in the linear prediction analysis of speech are summarized as follows: (1) it is more effective due to the optimization of the kriging equations taking into account the biased condition; (2) it is more flexible by allowing different biased values for the fitting of the signal spectrum, and therefore may provide a means for adaptive LPC; (3) it can give a good estimate of the number of poles used in the LPC by means of the theoretical semi-variogram.

Linear prediction analysis of speech based on set-membership theory

When computing linear prediction (LP) parameters of speech, large numbers of data are uninformative in a certain set-theoretic sense, and the expense of updating the estimates at these times can be avoided. “Set-membership” (SM) identification is formulated as a weighted recursive covariance LP problem with a special criterion for dynamic weight determination. An algorithm is developed which can be implemented on a systolic processor if desired, but which retains a simple interpretation as a specially weighted convariance LP method. The algorithm is applied to identification of the LP parameters of real speech data, and a number of practical issues are discussed. The potential for an adaptive strategy and other open research questions generated by the experimental work are discussed in the conclusions.

Advantages of a Givens rotation approach to temporally recursive linear prediction analysis of speech

Some useful features of a temporally recursive Givens rotation method in linear prediction analysis of speech and other signals are considered. Among them are convenient computation of the Itakura distance, windowed and exponentially adaptive estimation, lower rank estimation based on singular value decomposition, and set-membership identification. >

Some properties of the Givens' reduction applied to linear prediction analysis of speech.

Described here are some physical interpretations for the computational process of the Givens' reduction, an efficient algorithm for obtaining the least-squares solution of a set of over-determined linear equations. These physical interpretations are brought into view by applying the Givens' reduction to the linear predictive analysis of speech, and would not have suggested themselves if the Givens' reduction were considered only in connection with the solution of an ordinary set of over-determined linear equations. The attractive advantages found in the computational process of the Givens' reduction are the following:(1) The forward and the backward prediction errors for each data sample are automatically obtained in the working matrix of the augmented Givens' reduction.(2) The direct time-update recursion for the K-parameter, the reflection coefficients in linear prediction analysis, is implicitly performed sample by sample.(3) An index representing the rank of the covariance matrix of the input data sequence is available in the working vector. These advantages come out from the physical interpretation of the corresponding parts of the working areas for the Givens' reduction. The behavior of these values in the working areas are empirically confirmed by actual analysis of speech.

Linear prediction analysis of speech based on a pole-zero representation.

Speech analysis and synthesis by linear prediction is based on the assumption that the short-time spectral envelope of speech can be represented by a number of poles. An all-pole representation does not provide an accurate description of speech spectra, particularly for nasals and nasalized sounds. This paper presents a method for characterizing speech in terms of the parameters of a pole-zero model. In this method, an impulse response representing the composite filtering action of the glottal wave, the vocal tract, the radiation, and the speech recording system is first constructed from the speech signal. This impulse response is obtained by performing several stages of all-pole LPC analysis. The pole-zero parameters are determined from the impulse response by solving a set of simultaneous linear equations. The method, being noniterative, is very suitable for automatic analysis of speech. The method has been applied to real speech data and the results show that the speech spectra derived from the pole-zero model agree very closely with the actual spectra derived by direct Fourier analysis.

All-pole modeling of degraded speech

This paper considers the estimation of speech parameters in an all-pole model when the speech has been degraded by additive background noise. The procedure, based on maximum a posteriori (MAP) estimation techniques is first developed in the absence of noise and related to linear prediction analysis of speech. The modification in the presence of background noise is shown to be nonlinear. Two suboptimal procedures are suggested which have linear iterative implementations. A preliminary illustration and discussion based both on a synthetic example and real speech data are given.

On determining partial correlation coefficients by the covariance method of linear prediction

Partial correlation coefficients of speech samples (sometimes called reflection coefficients) are an important set of parameters in linear prediction analysis of speech and have proved to be useful for many practical applications, such as speech encoding, speech synthesis-by-rule, etc. So far, it has been thought [Gibson, IEEE Trans. ASSP, p. 93 (Feb. 1977)] that the covariance method does not provide a direct method of computing the partial correlations unless the covariance matrix is a Toeplitz matrix. Indirect computation of partial correlations from the predictor coefficients have obvious disadvantages. It will be shown that the partial correlation coefficients are indeed obtained as a set of intermediate parameters by the covariance method even when the covariance matrix is non-Toeplitz. These intermediate parameters are obtained as a solution of a set of linear equations through Choleski decomposition of the covariance matrix and, after appropriate normalization, are transformed to the partial correlation coefficients. The above method of determining partial correlations has been extensively tested with both real and synthetic speech data and has been found to work satisfactorily with analysis intervals as short as one pitch period.

Linear prediction analysis of speech based on pole-zero representation

Speech analysis and synthesis by linear prediction is based on the assumption that the short-time spectral envelope of speech can be represented by a number of poles. An all-pole representation does not provide an accurate description of speech spectra, particularly for nasal and nasalized sounds. In this paper, a method is presented for characterizing speech in terms of the parameters of a pole-zero model. The basic approach used in the method is similar to one used for the all-pole case: the parameters of the pole-zero model are obtained by minimizing the mean-squared prediction error over the analysis interval. It is shown that, although the equations resulting from the minimization are nonlinear, both poles and zeros can be determined without using an iteractive procedure. The method has been applied to real speech data and the results show that the speech spectra derived from the pole-zero model agree very closely with the actual spectra derived by direct FFT analysis. Audible distortion due to inadequate treatment of zeros in the all-pole case disappear with pole-zero representation.

A spectral-flatness measure for studying the autocorrelation method of linear prediction of speech analysis

The purpose of this paper is to introduce a spectral-flatness measure into the study of linear prediction analysis of speech. A spectral-flatness measure is introduced to give a quantitative measure of whiteness, of a spectrum. It is shown that maximizing the spectral flatness of an inverse filter output or linear predictor error is equivalent to the autocorrelation method of linear prediction. Theoretical properties of the flatness measure are derived, and compared with experimental results. It is shown that possible ill-conditioning of the analysis problem is directly related to the spectral-flatness measure and that prewhitening by a simple first-order linear predictor to increase spectral flatness can greatly reduce the amount of ill-conditioning.