The LaMIT project database
A lexical access database for Italian will be conceived and recorded. This database will be structured along similar principles underlying the MIT English database, i.e. about 100 sentences pronounced by 4 speakers (two males and two females). The database will be designed so to reflect the typical frequency of occurrence of the different phonemes in Italian, as suggested by a recent study (
Arango et al., 2020) that provides updated values. Attention will also be given to reflecting crucial characteristics of the phonology of the language, in particular consonant gemination which is word contrastive in Italian. In Italian gemination may also be syntactic; This phenomenon, known as “raddoppiamento sintattico”, is shared by Italian with very few other languages such as Finnish and in some way Maltese for Italian and Sicilian imported words.
Once recorded and digitalized, the database will be labeled with the cues to the hierarchically organized features of Italian phonemes. To achieve this goal, a specific software will be developed to create a well-structured and easily accessible labeled database containing exhaustive information on the predicted cues to landmarks and other features. Manual labeling will then be performed, that is, for each sentence an experienced speech scientist will verify the actual presence of the cues by examining the properties of the signal and its spectrum. The manual labelling process will also address the gemination phenomenon, not only in its predictable lexical and syntactic forms, but also in the so-called expressive gemination, typical of speaking and dialectal styles that may not be predicted but may manifest in the spoken sentences and its acoustics.
A fully-integrated labeled database will allow to perform extensive acoustic analyses on Italian speech data, the results of which will be used to derive the acoustic cues to landmarks and features. This phase will also be crucial to understand whether landmarks vs. other features are language independent. The adopted method will consist in developing and testing a landmark detection recognizer on both the Italian and English databases; This will shed light on the robustness of the adopted approach across languages.
While designing the speech recognition system, i.e. the complex association of modules dedicated to the different functions of the system itself - detection of landmarks, detection of specific features, and ultimately word recognition detection of landmarks, detection of specific features, and ultimately word recognition - we will set the hypotheses for a principled introduction of inferred features, the goal here being to better understand and model the physical system at hand, and not to bypass the lack of knowledge by introducing brute-force statistical approaches that may blindly neglect underlying articulatory and acoustic hidden regularities.