Effects of Room Reverberation on Speaker Identification

Accuracy in speaker recognition systems may degrade if channel conditions during the training and testing stages are significantly different. Such channels may include different microphones, telephone and mobile handsets, speech coders, and VoIP. Many compensation techniques have been proposed which seek to minimize the channel mismatch condition thereby improving accuracy rates in these systems. More recently, the acoustic channel and its effect on speaker identification (SI) have been investigated and it has been shown that when using clean training signals and reverberated test signals, a loss in accuracy results. In this paper, we improve upon a proposed method to compensate for this acoustic channel mismatch by utilizing a more accurate room reverberation model during the training stage. This model allows us to pre-distort (reverberate) clean training signals in order to approximate the expected reverberation present in test signals. By utilizing a set of reverberated training models for each speaker, SI accuracies can be improved.