Multi-frame rate based multiple-model training for robust speaker identification of disguised voice

Speaker identification systems are prone to attack when voice disguise is adopted by the user. To address this issue, our paper studies the effect of using different frame rates on the accuracy of the speaker identification system for disguised voice. In addition, a multi-frame rate based multiple-model training method is proposed. The experimental results show the superior performance of the proposed method compared to the commonly used single frame rate method for three types of disguised voice, namely, synchronous, fast and repetitive synchronous imitation taken from the CHAINS corpus.