Automatic Speaker and Language Identification

This thesis deals with the problem of automatic language identification (LID) and automatic speaker identification (SID) given the speech signal as input. Both researches have received renewed interest due to heightened homeland security awareness, e.g. in the use of speaker’s voice print for biometric identification, language identification for the classification of speech archives and call-scanning. In addition, technological progress on speaker and language identification (SLID) research will also enhance the robustness of automatic speech recognition. The technologies used in SLID applications are closely related as both apply statistical models to identify speaker and the language identity. The key to solve the SLID problem is the detection and exploitation of differences between languages and speakers. The speaker and languages differ from one another along many dimensions, these include, phoneme inventory, phonotactic, prosodics, syllable structure, lexical words and grammar. These differences must be identified and extracted from raw speech data to successfully attack the SLID issues. While the progress of speaker and language identification research in the last few years has been heartening, significant improvements is still required for real-world environment applications as available data from these environments is often noisy and/or has short duration. Although human can normally successfully apply SLID using these data, current systems are unable to perform as robustly. This indicates that current systems have not fully exploited all available information in the recording and improvements can be made. The first year of the research focused on general literature study and investigating different approaches in language identification techniques. The research is focused on investigating the discriminative features and their fusion in language identification. We