Text-Independent Robust Speaker Identification

Abstract

As an underlying topic in speaker recognition, speaker identification aims to identify the speaker in a speech sample. Although identification of speaker is a complex problem, with the development of signal processing algorithms the problem has been simplified and the system performs better if the training and testing conditions are identical. However, the real world environment like background noise, room reverberation, crosstalk, etc., degrades the system performance. Achieving robustness in speaker identification has become a major concern ubiquitously. There are number of existing approaches to this problem such as proposing a robust feature set, introducing noise to models created by clean speech and using methods for speech enhancement to restore the clean speech characteristics. This dissertation aims to address the robustness problem in identification of speakers by proposing a new feature set and then introducing noise in the clean speech models. A new methodology is proposed in which root compression based features are extracted and the order of root is increased for analysing the apt value of root that gives the best identification results. Another methodology extracts the useful information from these feature set, making another set of features which performs substantially better than the conventional speaker features under the influence of noise. The system has been tested under a wide range of signal-tonoise ratio (SNR). The two methods of classification has been used namely, gaussian mixture models and support vector machine. To justify the results of identification, the proposed methodology have been verified on NIST-2003 and Vox-forge 2015 database in presence of noise. The analysis has been done with improving the order of root in the feature compression stage and the evaluation results have shown that features extracted by taking the square root have outperformed others in a noise dominant system while the cube root compression based features have shown best identification accuracy in case of signal dominant system