Automatic Recognition of Italian I-Set by Recurrent Neural Networks

In order to prove the potential power of "learning by examples" paradigm for problems of Automatic Speech Recognition, an experiment is described, regarding an extremely difficult Italian phonetic recognition problem: the automatic discrimination of the so called Italian i-set: /bi/, /tSi/, /di/, /dZi/, /i/, /pi/, /ti/, /vi/ plus other two i-like stimuli /Li/, /si/. Auditory Modeling is used as front-end digital signal processing. Semi-automatic Multi-Level segmentation is applied to input speech stimuli. Recurrent Neural Networks trained by Extended Back Propagation for Sequences constitute the global recognition framework.. The achieved speaker independent mean recognition rate is around 65% which, given the effective difficulty of the present task, can be considered quite acceptable and promising.