Phoneme recognition via coupling landmark ISOMAP and Random Forests

In this paper we proposed the usage of a non-linear dimensionality reduction technique for the task of phoneme recognition. The classic ISOMAP (Tenenbaum et al) is computationally expensive, hence for large datasets is an ihnerently prohibitive, albeit we can approximate the quality of mapping via using landmark ISOMAP, which essentially is an approximation of ISOMAP with theoretical guarantees. The main gist of our solution is to increase the dimensionality of feature vector and then to project to a lower dimensional manifold. By performing this step we try to encapsulate non-linearities that exist in feature space that cannot be otherwise revealed.Classification of phonemes is performed via Random Forests which is computationally light and has strong probabilistic background. We accompany our work with experiments over a subset of phonemes drawn from TIMIT Database.