Bi-level Speaker Identification System: Impact on Short Utterances
Author Name : Leila Beltaifa-Zouari, Rania Chakroun
ABSTRACT: Despite the recent advances in the field of speaker recognition, there are still many problems for which solutions have to be found. In particular, recognizing a speaker’s identity when only a little amount of speech data is available remains a key consideration since many real world applications have access to speech data having limited duration. In this paper, we present a new approach based on additional information detected from the speech signal to improve the task of automatic speaker identification. In doing so, we highlight how the detection of the speaker’s dialect can be explored to address the research problem related to Short Utterance Speaker Recognition (SUSR). Our study suggests that the automatic detection of the dialect of the speaker can be useful to deal a new approach for speaker identification system when training data are limited and test utterances are very short. We perform the proposed approach in two steps: at the beginning the dialect of the speaker is automatically detected then the identity of the speaker is automatically recognized. We achieved remarkable results using TIMIT database and Support Vector Machines technique. In this context, the proposed approach presents an efficient solution for the constraints related to the memory and computational resource limitation in realistic applications, and hence makes possible the use of large datasets containing many speakers.