An open and free speech corpus for speaker recognition: The FSCSR speech corpus

Speech corpora are considered as the basic requirement for the development and evaluation of speaker recognition systems. However, the vast majority of speech corpora intended for speaker recognition research is not free for researchers, and the remaining freely available speech databases are not sufficiently large for the development and evaluation of robust speaker recognition systems. This lack of free and large speech corpora represents a major obstacle for promoting research on speaker recognition systems, especially in less-resourced laboratories. To overcome this problem, we have developed a free and large speech corpus of more than 530 speakers from several languages. A part of the proposed speech corpus is collected and recorded under our supervision, and the remaining part is collected from the Voxforge website using a new semi-automatic approach. In this paper, the detailed description of the proposed speech corpus and the complete procedure of data collection are presented. The results of the preliminary experiments conducted on the proposed speech corpus, using speaker recognition state-of-art approaches (MFCC for features extraction, GMM-UBM for speaker modeling), are also presented.