Estimating Tongue-Palate Contact Patterns from the Speech Signal

Electropalatography (EPG) is a technique that determines the contact patterns between the tongue and the hard palate during speech. In one of its most common forms, it utilizes an artificial palate with 62 silver electrodes embedded in its tongue-facing surface. At small regular time intervals it is recorded whether a specific electrode is contacted or not by the tongue, leading to tongue-palate contact patterns. EPG is nowadays a relatively well-estabilshed tool in phonetic research, in the clinical treatment of people with articulation difficulties or cleft palate, and also in the teaching of second languages. Still, the derivation of EPG data is a rather expensive and difficult process. What is suggested herein is that a means of estimating EPG patterns directly from the acoustic speech signal with no need of any special equipment would be of great value to speech pathologists and phoneticians alike. This paper presents work towards finding a mapping between acoustic parameters, namely the Mel Frequency Cepstral Coefficients, derived directly from the speech signal, and the corresponding EPG patterns. It may be regarded as a special case of a more general problem called acoustic-toarticulatory inversion, or speech inversion, which refers to finding mappings between the speech signal and some kind of articulatory parameters. One of the main motives for this research field is that the additional articulatory information could be used to improve the performance of current speech recognition systems. EPG patterns could also be used in such a context. For a solution of the problem described, we investigate the utilization of Support Vector Machines, a relatively new and very promising supervised learning technique, of which not many applications have yet appeared in the speech processing field. The source of the data we use is the MOCHA database, which is well documented and publicly available via the Web, thus allowing for comparisons of other researcher’s results to ours.