Modeling the process that a listener actuates in deriving words intended by a speaker, requires setting a hypothesis on how lexical items are stored in memory. This project proposes to test Stevens's hypothesis of Lexical Access (Stevens, 2002), that is, that words are stored as a hierarchical arrangement of distinctive features, by extending the analysis to Italian. Exploring a new language will provide insight into whether Steven's approach has universal application across languages, with relevant implication for the understanding of how human brain recognizes speech. A principled introduction into the model of the concept of inference is also proposed. While developing a complete Italian recognizer is my long-term goal, this project forms its founding block, and combines my two professional activities in speech and impulse ultrawideband radio modeling.
In May 2019, I was named a 2019-2020 fellow at the Radcliffe Institute for Advanced Study at Harvard University, joining more than 50 women and men in the incoming fellowship class as they pursue work across the sciences, social sciences, humanities, and arts. As the 2019-2020 William Bentinck-Smith Fellow, I will pursue the project in a community dedicated to exploration and inquiry at Harvard's Institute for Advanced Study.
The expected outcomes with significant impact of the work are:
1. The Lexical Access model has been so far only applied to American-English. Its application to a different language may lead to an understanding of the underlying universal language-independent nature. In particular, the model postulates the relevance of specific acoustic discontinuities - the landmarks - that correspond to categorized distinctive features (i.e. manner features). An important outcome will be the understanding of language-independent properties of acoustic landmarks.
2. The introduction of the concept of inference in the model will be principled. The danger would be to bypass the lack of knowledge by introducing brute-force statistical approaches that blindly neglect underlying articulatory and acoustic hidden regularities - as currently adopted in some speech recognizers that use machine learning and neural networks - and by doing so contradict the founding principles of the Lexical access model. Rather, the goal here is to better understand and model the physical system at hand, beyond attaining, in a short period of time, the improvement of recognition scores, if this is not ultimately obtained by improved modeling. Which parts of the system are suitable for the inference concept, and for which reasons? How does the introduction of the inference concept impact the decision rules that lead to the recognition of words? A major outcome may show that the findings are applicable in turn to other languages, and in particular to American-English for which the Lexical Access model was originally conceived.
3. To the best of my knowledge there exists no system of such kind for Italian. The development of an Italian speech recognizer based on detection of landmarks, and the possibility to provide open access to data and algorithms with the creation of a publicly accessible website of the project, may lead to establishing a reference record for the speech community in Italy and abroad, based on the tremendous impact the project would provide in terms of supporting evidence for universal strategies of speech perception.