Influence of Noise and Voice Activity Detection on Speaker Verification.

The scope of this paper is to check influence of voice activity detection VAD procedure and its accuracy on speaker verification error rates. It is shown that for speech of high quality, it is absolutely necessary to remove silence from the signal as the errors increase radically. It is better to remove more than less from the signal as the equal error rate EER is the worst for the original speech with silence. Additionally influence of white noise, which was added to speech utterances, was examined. Presented results show that in order to achieve highly reliable speaker verification system it must be insensitive to low quality of speech, since noise is the most important factor responsible for high error rates.