Automatic English phoneme recognition from articulatory data generated by EPG systems with grid and anatomical layout of contact sensors

The aim of the study was to conduct automatic phoneme identification from articulatory data that accompanied the production of these phonemes in continuous speech. The articulatory data were obtained from 2 electropalatographic systems, Palatometer by Complete Speech and Linguagraph by Rose-Medical. Palatometer was used with the artificial palate containing 124 contact sensors in a grid layout, including 2 sensors monitoring the lip contact. The palate included a vacuum-thermoformed flexible printed circuit. Linguagraph was used with the acrylic artificial palate designed and developed for the purpose of this study, containing 62 electrodes in anatomical layout. Palatometer was used by one native of General American and Linguagraph by one native of General British, each reading 140 phonetically balanced sentences that included Harvard Sentences and TIMIT prompts. The EPG data were parametrised into dimensionality reduction indexes, which were analysed by means of linear discriminant analysis and a probabilistic neural network. The results of classifications are discussed.