Statistical identification of articulation

We present a statistical technique for identifying critical, dependent and redundant roles played by the articulators during production of English phonemes using articulatory (EMA) data. It identifies a list of critical articulators for each phone based on changes in the distribution of articulator positions. The effect of critical articulation on dependent articulators is derived from inter-articulator correlation. Articulators unaffected or not correlated with the critical articulators are regarded as redundant. The technique was implemented on 1D and 2D distributions of midsagittal articulator coordinates, and the results of this data-driven approach are analyzed in comparison with the phonetic descriptions from the IPA chart. The results using the proposed method gave a closer fit to measured data than those estimated from IPA information alone and highlighted significant factors in the phoneme-tophone transformation. The proposed algorithm was evaluated against an exhaustive search of critical articulators, and found to be as effective as the exhaustive search in modeling phone distributions with the added advantage of faster execution times. The efficiency of the approach in generating a parsimonious yet accurate representation of the observed articulatory constraints is described, and its potential for applications in speech science and technology discussed.