Received September 28,
2011
Abstract. Text-independent speaker
identification and verification (a.k.a. open-set speaker identification) is one
of the most important problems in design of automatic speech recognition
systems. This paper describes the method based on sparse representations that
enhances recognition accuracy. The effectiveness of the proposed approach is
demonstrated using two independent evaluations with phone-quality speech
signals. It is shown that applying sparse representations in the open-set
speaker identification problem reduces the equal error rate by more than half,
while considerably increasing the identification rate.
Keywords:
automatic speaker
identification and verification, Mel-Frequency Cepstral Coefficients (MFCC),
sparse representations, GMM supervector, phone-quality audio database.