Speaker identification with whispered speech mode using MFCC: Challenges to whispered speech identification

Whispered mode of speech is preferred by people for confidential communication. To identify speaker from his whispered mode of utterances is challenging compared to neutral speech as the speaker identification rate is found to be reducing in the whispered case. This paper uses the MFCC (Mel-Frequency Cepstrum Coefficients) for feature extraction and VQ (Vector Quantization) for matching to identify speaker. Here, closed set of 35 speaker's database is generated to test the system: both whispered and neutral utterances of all speakers are recorded. When trained with neutral speech sample and test query was also neutral, identification is 91.4%. At the same time for neutral speech trained system when the test sample was whispered, identification rate reaches as low as 68.56%. The identification of unknown speaker is based on the Euclidean distance. The GUI created in MATLAB is used for displaying the result of speaker identification. The systematic analysis is made to find the reasons of deteriorating performance in the whispered case and probable remedies are suggested.