Influence of the Number of Principal Components used to the Automatic Speaker Recognition Accuracy

摘要

It is widely accepted that voice is a behavioral trait that may be used as a biometric characteristic. Systems intended to exploit this property of voice must be able to make a credible representation of the observed voice, and equipped with reliable procedures for automatic speaker recognition [1]. Feature extraction is one of the key components of automatic speaker recognition systems. Mel-Frequency Cepstral Coefficients (MFCCs) and their first and second derivatives are often used as a feature set. The observed speech signals are characterized by their predictive nature. MFCCs may be considered as a direct consequence of discretization of the spectrum envelope, and their derivatives as linear combinations of any two adjacent MFCCs in the infinitesimal time. Therefore, it follows that these features are mutually redundant [2]. On the other hand, the basic problem of achieving a high level of reliability of speaker recognition lies in the fact that training and testing conditions differ. Since each feature is susceptible to environmental interference, the increased complexity of a speaker model (i.e., the increased dimensionality of feature vectors) implies an increased level of noise. From the methodological point of view, to prevent such a level of noise, it is necessary to focus the recognizer on what is important in the observed recognition object. In other words, reduction of the model complexity would reduce the accumulated noise level. These observations motivate possible efficiency improvements of a speaker recognizer. Transformation techniques that are usually applied for dimensionality reduction include Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Nonlinear Discriminant Analysis (NLDA) [3]. All these techniques are intended to train the recognizer to focus on those elements that are essential for the observed data set. Recent work indicates that PCA is a prominent technique for dimension reduction. Conventional methods for PCA based on the full data covariance matrix require a large amount of training data [4]. In order to reduce the complexity of these methods where the eigenvector matrix of each speaker is calculated, methods for PCA that are performed on all the training data [4, 5] (and this paper), or on locally clustered data [6, 7] are introduced. In the domain of speaker recognition, PCA is often applied in Gaussian Mixture Models (GMMs) [4, 7]. In the next section, we introduce the reference environment and the modus of speaker modeling. Then, we consider the feature transformation and training of the model. Finally, we discuss the results of automatic speaker recognition.