A new Italian dataset of parallel acoustic and articulatory data

In this paper we introduce a new Italian dataset consisting of simultaneous recordings of continuous speech and trajectories of important vocal tract articulators (i.e. tongue, lips, incisors) tracked by Electromagnetic Articulography (EMA). It includes more than 500 sentences uttered in citation condition by three speakers, one male (cnz) and two females (lls, olm), for approximately 2 hours of speech material. Such dataset has been designed to be large enough and phonetically balanced so as to be used in speech applications (e.g. speech recognition systems). We then test our speaker-dependent articulatory DeepNeural-Network Hidden-Markov-Model (DNN-HMM) phone recognizer on the set of data recorded from the cnz speaker. We show that phone recognition results are comparable to the ones that we previously obtained using two well-known British-English datasets with EMA data of equivalent vocal tract articulators. That suggests that the new set of data is a equally useful and coherent resource. The dataset is the session 1 of a larger Italian corpus, called Multi-SPeaKing-style-Articulatory (MSPKA) corpus, including parallel audio and articulatory data in diverse speaking styles (e.g. read, hyperarticulated and hypoarticulated speech). It is freely available at http://www.mspkacorpus.it for research purposes. In the immediate future the whole corpus will be released.