Text-Independent Speaker Verification using Optimized Linear Combination of Local MFCC Features

S. Sakai and K. Kameyama (Japan)


Text-independent speaker verification, biometrid authentication, adaptive feature weighting, inter-frame feature.


In recent years, studies of speaker verification have been conducted as a means for biometric person authentication. However, because of the overall verification performance, only few actual implementations exist. This paper focuses on the text-independent speaker verification system. We propose an effective method for speaker verification by adaptive weighting of local Mel Frequency Cepstrum Coefficient (MFCC) features. For a given set of registered persons, optimal linear weightings of multiple speech frames are searched based on the likelihood ratio error, generalizing the scheme of the conventional use of Δ parameters [1]. It was observed that using the proposed adaptive parameters, superior verification performance was achieved compared with the cases using conventional features.

Important Links:

Go Back