Spoken-digit recognition using self-organizing maps with perceptual pre-processing

Abstract

One of the current challenges in Pattern Recognition is the treatment of Speech, the complexity of this task being due fundamentally to the great statistical variability of the speakers, and to the temporal structure of the speech signal in itself. Throughout the current bibliography multiple discriminative methods and coding algorithms can be found claiming slight advances in the recognition rates, which may be considered important advances, as the field is reaching a verge difficult to move over. In this experiment a representation in bidimensional selforganizing maps of the decimal digits spoken in English (from one to nine is carried out. This representation has been checked taking data from the TIMIT database, starting from a previous code based in Perceptual Linear Prediction coefficients (PLP). Subsequently, a heuristic algorithm for recognition has been defined. The application of this algorithm to both a training data set and a test data set produces acceptable recognition rates, even for low-dimension maps with the benefit of the reduction in the computational costs. The basic methodology used and the mentioned results are presented and discussed.