Phonetic analysis of a computational model for vocabulary acquisition from auditory inputs

An overview of the NMF (non-negative matrix factorization) approach for word acquisition from auditory inputs is given. The method can link acoustic realizations of spoken words with information observed in other modalities. The method is applied to the acquisition of a small set of keywords embedded in carrier sentences. A phonetic description is linked to the learned representations which are otherwise difficult to interpret. Through this link, the phonetic similarity between the learned acoustic representations and lexical items is displayed and interpreted.