Speaker Verification In Multi-Channel Condition

Automatic speaker recognition, also known as voiceprint recognition, is an important branch of speech signal processing, and is also one of the most popular research fields as one kind of biometrics. The area of speaker recognition is concerned with extracting the identity of the person speaking the utterance. Depending on the application, the general area of speaker recognition is divided into two specific tasks:speaker verification and speaker identification. In this paper, the basic principle of speaker recognition will be introduced, and the research history and status quo is also mentioned. While explaining the dominant method used in text-independent speaker verification task, namely GMM-UBM, in detail, this paper also discusses the strategy of using support vector machine (SVM) classifier to do the speaker verification task, both the theory and system design will be included. To the factor of channel mismatch that impacts system performance significantly, this paper discusses various channel compensation techniques in feature domain, model domain and score domain. I also give my own understanding and improvement opinions to some of these techniques.In my study, the prevalent, mutual and open-source speech recognition toolkit developed by Cambridge University is used to build my speaker verification system. This system compared the performances of several different techniques used in feature domain and model domain, etc. In my final system, the PLP static feature parameter appended by its first, second and third order parameter is adopted, and the various techniques, such as RASTA filtering, feature mapping, feature transformation, model adaptation, score normalization, are used to optimize my system. As the experimental results shown, my final recognition system performs well and its performance approaches the state-of-the-art level in this field, according to the results announced by NIST organization for current few years.In this paper, my research is mainly focused on the text-independent speaker verification system under telephonic channels. But it is worthwhile to point out that many techniques and methods are valuable for reference and application by other speaker verification tasks, speaker identification tasks and even speech recognition tasks.