Investigation of MFCC feature representation for classification of spoken letters using Multi-Layer Perceptrons (MLP)

In this paper, the Mel-Frequency Cepstral Coefficient (MFCC) is demonstrated as an effective feature representation method for spoken letters recognition. The Multi-Layer Perceptron (MLP) was used as a classifier to discriminate between two spoken letters - `A' and `S'. The dataset consists of 72 samples (35 and 37 samples of spoken letters `A' and `S', respectively). The samples were represented using the Mel Frequency Cepstral Coefficients (MFCC). Several experiments were conducted to determine the optimal network parameters to yield the best classification results. The results indicate that the optimal network structure was with 2 hidden units, which yielded classification accuracy of 100% (training) and 93% (testing).