Speaker similarity score based fast phoneme classification by using neighborhood components analysis

K-nearest neighbor (k-NN) classifier can learn non-linear decision surface and requires only one hyperparameter (i.e. value of “k”) for training. The classification accuracy improves as we increase the amount of training data. With an increase in the amount of training data, computational and memory requirements also increases as it has to store and search through the entire training data for classification of one test point. In this paper we investigate the computational and memory cost of speaker similarity score algorithm for phoneme classification by doing dimensionality reduction based on neighborhood components analysis. Our speaker similarity score algorithm uses k-NN for learning a speaker similarity score of a target speaker and uses this score for weighted k-NN phoneme classification. By using neighborhood components analysis we obtained a significant reduction in computational and memory requirements at the expense of a small phoneme classification performance loss. Experiments on TIMIT dataset shows 56% reduction in computational cost as we reduced the dimensions of our feature space from 50 to 22.