Automatic Voice Recognition System based on Multiple Support Vector Machines and Mel-Frequency Cepstral Coefficients

The problem of speech recognition could be considered a classification problem and modeled as such, where one wants to get the best degree of separability between classes representing the voice. In order to apply that concept to build an automated speech recognition system capable of identifying the speaker, many techniques using artificial intelligence and general classification have been developed, which lead to this work. Here is proposed a voice recognition method capable of recognize keywords in brazilian portuguese for biometric purpose using multiple Support Vector Machines, which builds a hyperplane that separates Mel Frequency Cepstral Coefficients, the MFCC's, for later classification of new data. With a small dataset the system was able to correctly identify the speaker in all cases, having great precision on the task. The machines are based on the Radial Basis Function kernel, the RBF, but were tested with severe different kernels, having also a good precision with the linear one.