Investigation of architecture for verification and recognition of speaker
Weber Campos, Alexandre Magno Andrade Maciel, Edson Damasceno Carvalho · 2008
The growing need to identify individuals becomes increasingly evident today, and biometrics are presented as the most safe and effective to perform this task. Among the various patterns that can be used in a biometric system, this work is dedicated to the study of the voice biometrics. Thus, a study of the area of voice processing and its applications were made and an architecture to systems of recognition and verification of a speaker was proposed. To that end, it was necessary to acquire a basic human voice to be built. The feature extractor MFCC (Mel Frequency cepstral coefficients) was used for the purpose of obtaining information that best representing each speaker. From the data obtained, was conducted a detailed study of the database (through statistical methods) and performance of Neural Networks MLP (Multilayer Perceptron) and SVM (Support Vectors Machines) were investigated to recognize the speakers. Besides In addition, the Vector Quantization also been applied to test speakers, showing satisfactory results.