An adaptative thresholding SOM-wavelet packet model to improve the phonemic recognition rate

This approach aims to combine the Kohonen map abbreviates in SOM (Self Organizing Map), as a powerful tool for recognition, with the binary technique of signal processing by wavelet packets for noise removing. The obtained model can be applied successfully to solve the phonemic recognition problem, pronounced by multi speaker in varying environmental conditions. The adopted method involves applying a suitable thresholding per node, of the signal decomposition tree, as an extension of thresholding level method. The noise estimation is based on the spectral entropy computation using the histogram's coefficients intensity of wavelet transform for each node in the tree, rather than the estimation based on the median coefficients in each node directed by the median absolute deviation (MAD). To solve the problem of time-frequency discontinuity introduced by hard thresholding, we'll use another nonlinear law called logarithmic law. The reconstructed signal will be set in Mel cepstrum coefficients to form the acoustic vector admission into a phonemic recognition map SOM.