Verso un sistema di riconoscimento automatico del parlato tramite immagini ultrasoniche

We provide a proposal for a silent speech interface based on Ultrasound Tongue Imaging (UTI) data. The model we present appears to be the first proposal for a silent speech interface based on UTI for the Italian language. The model is fed with ultrasonic data only (e.g. lips movements are not included in the model) and the extraction of information from ultrasound images is accomplished via a model of artificial neural network instead of standard hidden Markov Model. In the paper, we illustrate the phases of acquisition, filtering and modeling of articulatory data. Then, we describe the training of the neural network, with particular emphasis on the problems associated with the process of identification and extraction of salient information. Eventually, we report the success rate for the model when recognizing Italian real words.