Mitigate the Reverberant Effects on Speaker Recognition via Multi-training

Abstract

Speaker recognition techniques have been developed into a relatively mature status over the past few decades through continuous research and development work. Existing methods typically use robust features extracted from clean speech signals, and therefore in idealized conditions can achieve very high recognition accuracy. For critical applications, such as security forensics robustness and reliability of the system is crucial. The reverberation condition can be represented by two main parameters namely Reverberation Time (RT) and Direct to Reverberation Ratio (DRR) (which represent the distance of the microphone to the source). This paper presents an efficient method to mitigating or at least alleviates the impacts of reverberation upon speaker verification. Multi-condition training approaches are investigated to alleviate such detrimental effects. Three multi-condition training methods are then investigated to mitigate such detrimental effects. The first uses matched train/test speaker models based on estimated reverberation time (RT) values. The second utilizes two-condition training where clean and reverberant models are used. Lastly, a four-condition training setup is proposed and conducted to improve the system performance. The utilized data set building, for SV experiments, training, and speech test material are obtained from the University of Salford Anechoic chamber database (SALU-AC). Experimental results show the first and the last types of multi-condition training providing significant gains in performance relative to the baseline.