A Novel Automatic Voice Recognition System using a Graph-Based Clustering Algorithm
Abstract – Weconsiderhereanovel automatic unsupervised speech-dependent speaker recognition technique. The proposed approach clusters a set of speech sequences in voice classes corresponding to the generating speakers, whose number is unknown. The speech feature vectors of these utterances are constructed by using the mel-cepstral analysis. Then, an automatic unsupervised classification is performed on the obtained feature vectors. A graph theory-based vocal feature vector clustering method is proposed for this task. It groups the vocal sequences in a proper number of speaker-classes. Experiments and method comparison illustrating the effectiveness of the proposed technique are also described in this research paper.
Keywords - Unsupervised Speech-Dependent Voice Recognition; Mel-Cepstral Analysis; Speech Feature Extraction; Graph Clustering; Normalized-Cut Algorithm.