Learning Electropalatograms from Acoustics

Electropalatography is a well established technique for recording information on the patterns of contact between the tongue and the hard palate during speech, leading to a stream of binary vectors called electropalatograms, consisting of elecropalatographic events - contacts or non-contacts between the tongue and the palate. A data-driven approach to mapping the speech signal onto electropalatographic information is presented. A combination of principal component analysis and support vector regression is used, yielding classification scores of more than 93% on individual electropalatographic events, for a single speaker. This may be viewed as a special case of the, well-known in the speech community, speech inversion problem which refers to inferring production parameters from the speech signal