A Comparison of Graph Construction and Learning Algorithms for Graph-Based Phonetic Classification

Graph-based semi-supervised learning (SSL) algorithms have been widely applied in large-scale machine learning. In this work, we show different graph-based SSL methods (modified adsorption, measure propagation, and prior-based measure propagation) and compare them to the standard label propagation algorithm on a phonetic classification task. In addition, we compare 4 different ways of constructing the phonetic graph: graph construction based on acoustic features vs. first-pass classifier outputs, in combination with either standard k-nearest neighbor search, or mutual k-nearest neighbor search. The best results are obtained with first-pass classifier outputs, mutual k-NN search, and prior-based measure propagation.