가정용 로봇을 위한 GMM-UBM 기반 화자 식별 기법 = A GMM-UBM-based speaker recognition method for home robots

다국어 초록 (Multilingual Abstract)

In this paper, we have proposed a GMM-UBM-based speaker recognition method using small-scale voice data toprovide personalized interactions for voice-controlled home robots. The proposed method extracts speakerfeatures, including MFCC(Mel-Frequency Cepstral Coefficient)s and first/second-order derivatives, from the inputvoice after preprocessing and applies them to the GMM-UBM model to recognize the speaker. Considering thehome robot environment, we have constructed a small-scale dataset consisting of three males and two femalesinclude short commands test dataset which were consist of each speech less than 1 second and independent of thetraining data to verified the sentence dependency. The proposed method has shown an accuracy of 98.40% and afast processing performance of 17.63ms per each sample. The proposed method in this paper enables highaccuracy and fast processing for short utterances without the need for large-scale training data, making iteffectively applicable to voice-controlled home robot systems.