HMM-based text-dependent speaker recognition with handset-channel recognition

In this paper, classical Gaussian Mixture Model and Hidden Markov Model based speaker recognition approaches are compared in a text dependent task. To compare two approaches under different handset-channel conditions, real-life scenario speaker recognition database is collected. Using the database, match-mismatch condition experiments are conducted. To improve speaker recognition performance under unknown test channel condition, channel recognition prior to speaker recognition is proposed. The accuracy of channel recognition and its effects on speaker recognition performance is investigated. It is observed that, speaker recognition performance converges to ideal match case with channel recognition while achieving %80-90 channel recognition accuracy.