I-vector based speaker gender recognition

Automatic gender recognition has been becoming very important in potential applications. Many state-of-the-art gender recognition approaches based on a variety of biometrics, such as face, body shape, voice, are proposed recently. Among them, relying on voice is suboptimal due to significant variations in pitch, emotion, and noise in real-world speech. Inspired from the speaker recognition approaches relying on i-vector presentation in NIST SRE, it´s believed that i-vector contains information about gender as a part of speaker´s characters, and works for speaker recognition as well as for gender recognition in complex environments. So, we apply the total variability space analysis to gender classification and propose i-vector based discrimination for speaker gender recognition. The results of experiments on TIMIT corpus and NUST603_2014 database show that the proposed i-vector based speaker gender recognition improves the performance up to 99.9%, and surpasses the pitch method and UBM-SVM baseline subsystems in term of accuracy comparatively.