Search

In this article, we present a bimodal speaker identification method, which integrates both acoustic and visual features and where the two audiovisual stream modalities are processed in parallel. We also propose a fusion technique that combines the two modalities to make the final recognition decision. Experiments are conducted on an audiovisual dataset containing the 28 Arabic syllables pronounced by ten speakers. Results show the importance of the visual information that is provided by Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT) in addition to the audio information corresponding to the Mel Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Predictive (PLP). Furthermore, some artificial neural networks such as Multilayer Perceptron (MLP) and Radial Basis Function (RBF) were investigated and tested successfully in this dataset by presenting good recognition performances with serial concatenation for the acoustic and visual vectors. [ABSTRACT FROM AUTHOR]