Tid och plats: Onsdagen 6 oktober kl. 16-18, rum D389.


No word is ever pronounced in the same way by even the same speaker. To parse and resolve the variation, there exists a variety of speech perception and psycholinguistic models which considerably vary in how words are represented in the mental lexicon, how much detail is stored, and how the speech signal is mapped on to the lexicon. Our model, the Featurally Underspecified Lexicon (FUL) assumes that the representation of words is phonologically sparse and only one morphophonemic variant is stored. The idea is that the privative sparse feature representation accounts for a number of asymmetries that are typical in synchronic and diachronic phonological systems. From the perceptual side, the speech signal is converted into phonological features by roughly defined acoustic parameters and the features are then matched against the lexical representation with a 3-way matching logic of "match, mismatch, no-mismatch" and a coherence metric is used to evaluate the 'strength' of the matching features. The talk will present and motivate these steps. We will first present a phonological sketch of the model along with psycholinguistic and neurolinguistic experimental evidence. We will then present our speech recognition system which is based on exactly the same principles.



Maria Koptjevskaja-Tamm och Tomas Riad