Transductive phoneme classification using local scaling and confidence
Matan Orbach, Koby Crammer · 2012
We apply a graph-based Transduction Algorithm with COnfidence named TACO to the task of phoneme classification. In recent work, TACO outperformed two state-of-the-art transductive learning algorithms on several natural language processing tasks. However, although TACO is a general-purpose algorithm, it has not yet been used for tasks in other domains, nor applied to graphs with millions of vertices. We show its effectiveness, as well as its scalability, by performing transductive phoneme classification on data from the TIMIT speech corpus. In addition, we experiment with two methods for graph construction, including local scaling, previously used for unsupervised clustering. Our results show that local scaling combined with TACO outperforms other combinations of graph construction methods and graph-based transductive algorithms.