Robust speaker verification in reverberant conditions using estimated acoustic parameters — A maximum likelihood estimation and training on the fly approach

摘要

Speaker recognition has been developed into a relatively mature state over the past few decades through continuous research and development work. Existing methods typically use the robust features extracted from noise and reverberation free speech signals, and therefore can achieve high recognition accuracy only in idealised conditions. Excessive reverberation as often occurs in many real world applications is known to compromise recognition performance. In this paper, a maximum likelihood estimation algorithm is proposed for blind-estimate reverberation time from speech signals submitted for verification. The estimates are used to choose matched acoustic impulse response or transfer function for the including in the retraining or fine tuning of the pattern recognition model on the fly. The training on the fly approach alleviates the detrimental impact of reverberation on authentication accuracy. Experimental results have shown significant improvement in system performance in terms of reduced equal error rate and detection error trade-off. This paper discusses the rationale, details the algorithm. And discusses the potential and limitations of this new method.