Comparison of Windowing Function on Feature Extraction Using MFCC for Speaker Identification

The speaker identification system is built by two main blocks; the first part is used to extract features from the input, while the second part is to classify the results from the features in the first part. Selection of the method to perform feature extraction is very important to obtain the optimal feature set. Mel-frequency cepstral coefficients (MFCC) is a feature extraction method that is used to convert the speaker's voice into coefficients as input for the classification process. There are several processes in MFCC, one of which is windowing. Windowing aims to reduce the discontinuous effect on the signal after the framing process. It is therefore important to use optimal windowing techniques so that the features of each sound are not wasted. This article highlights the use of several window functions such as hanning, hamming, bartlett, blackman, kaiser, and gaussian. The classification process proposed in this study is Artificial neural network (ANN). The data used amounted to 800 data from 16 speakers who were recorded directly. The data recorded for identification was the sound from the digits zero to nine (0-9) by each speaker. K-fold cross-validation was used as an evaluation of the classification model created to determine the combination with the best accuracy. The results shows that the use of 13 MFCC features with windowing hamming and gaussian with standard deviation values 72 obtains the best results. Both obtained an accuracy of 95%. This paper helps readers to gain insight in the field of speaker identification.