Speaker independent robust phoneme recognition using Higher-Order statistics and entropic-based features in adverse environments

In this work we present an algorithmic scheme for speaker independent robust phoneme recognition in noisy environments. The main modules of our proposal are composed by computing novel features that are based on Higher-order statistics, Rényi and Shannon's entropy and then by using Random Forests as classifier of the system. The main motivation is to combine ideas from different fields, that have been partially successful, in order to characterize, shed light and tackle the hardness of robust phoneme recognition. Our experiments were carried out over a subset of phonemes of the TIMIT database and the results show potential of the method in environments of varying degrees in the presence of noise and even competing and surpassing state of the art methodologies found in the literature.