Research On Robust Language Identification

Language Identification is the process of recognizing the language being spoken from a sample of speech with a computer, and it has a wide range of applications. This paper studies the key techniques of the Robust Language Identification. In this dissertation, we explore the possibilities to obtain high identification rate while enhancing the robustness of the LID system. They are explored in data obtaining, feature extraction, channel compensation in feature domain, supervector transforming and SVM kernel selecting. Several novel and efficiency algorithms are proposed. The main work and innovations obtained in this paper can be summarized as follows.(1) The GSV-SVM model needs large amount of training data, while several problems exist in the VOA linguistic data acquisition method commended by NIST, including great deal of speech in the background and non-target languages in the clean speech segment. In order to solve these problems, an improved VOA linguistic data acquisition method based on GSV-SVM and Computational Auditory Scene Analysis is proposed. First, the GSV-SVM LID system is used to distinguish the target language segments from the non-target ones. Then, The Computational Auditory Scene Analysis is utilized to detect the segments containing speech in the background. Finally, the experiments show that our approach obtains much cleaner linguistic data and outperforms the VOA method when applied to the LID system.(2) In current language identification system, the most commonly used MFCC feature parameters have not made the best of auditory characteristics and have weekly robustness in complex environments. An auditory-based robust feature extraction algorithm is proposed in this paper. The Sub-band energies of the extracted auditory features are calculated using a Gammachirp filter bank instead of the commonly used triangle filter bank. The compensation filter by data-driven analysis for each Sub-channel output is obtained by a constrained optimization process which jointly minimizes the environmental distortion as well as the distortion caused by the filter itself. Tests show that the new features outperform the MFCC feature in real environments.(3) In order to improve the robustness of the LID system, Factor Analysis technique is introduced to compensate the channel, and two improvements are made. Firstly, the Eigenchannel subspace estimation is used to replace the original noise subspace estimation method in supervector space. And with the channel factor estimated, the features are compensated using channel adaptation in the feature domain. Experiment results show that this method can work efficiently, and the identification rate is remarkably improved when the Factor Analysis technique is applied in the test in different channels.(4) GSV-SVM system is the state-of-the-art Language Identification framework. It uses Kullback-Leibler (KL) kernel to measure the distance between two GMM, but neglects the covariance information, and the dimension of the GSV is too large and limits the efficiency of the SVM training and identification. In order to eliminate these problems, we propose a LID system based on Bhattacharyya kernel and Hierarchical Heteroscedastic Linear Discriminant Analysis (HHLDA). The Bhattacharyya kernel ensures that both the mean value and covariance are used, and the HHLDA greatly reduces the dimension of the GSV, so that less data are needed for the training. The experiment results show that the proposed method has high identification rate and efficiency, and it outperforms existing methods.