Acoustic reduction and the roles of generalizations and exemplars in speech processing

Mirjam Ernestus

1: Radboud Universiteit Nijmegen 2: Max Planck Institute for Psycholinguistics

TP187: Usage-based and rule-based approaches to phonological variation

In spontaneous speech, a word like "yesterday" is often pronounced as "yesay" or appears in some other reduced form. Since reduced forms are so frequent, psycholinguistic models of speech processing should be able to account for how language users process them.

According to abstractionist models, the lexicon typically lists only one, unreduced, pronunciation for every word. These models explain the processing of pronunciation variation by assuming abstract rules, which compute the required pronunciation on the basis of the unreduced form during production, and deduce this unreduced form from the different possible pronunciations during comprehension. Exemplar-based models, in contrast, assume that language-users have stored all tokens of a word that they have ever heard. During production, speakers select one of these tokens for imitation. During comprehension, they match the perceived pronunciation with the most similar exemplar.

In order to investigate the production and comprehension process of reduced forms, we conducted several corpus studies and psycholinguistic experiments in Dutch. We found that neither a simple abstractionist nor a simple exemplar-based model can account for all data.

First, a word's degree of reduction is predictable to some extent on the basis of its phonetic structure, its phonetic context, its pragmatics, morphological structure, frequency of occurrence, predictability in the given sentential context, the number of previous mentions in the same conversation, and the speaker's gender and social class. Second, listeners appear to need at least phonetic and semantic / syntactic context for understanding reduced forms: Listeners only recognize the form "yesay" as "yesterday" if presented in its sentential context. This shows that abstractionist models have to assume very complex rules, taking into account many different types of information. These rules have to explain, for instance, how Dutch listeners know that "eik" is sometimes a reduced form of "eigenlijk" 'actually', rather than the word "eik" 'oak' itself, and never a reduced form of "ijselijk" 'hideous'. Exemplar-based models, in contrast, have to assume that tokens are stored together with their broad (socio-)linguistic context (as, e.g., in Polysp, Hawkins & Smith 2005). They face the challenge of explaining why all words show very similar patterns of reduction, even though the storage of exemplars allows for word-specific variation.

Third, speakers are typically not aware of the reduced forms occurring in their own language. Upon hearing a reduced form, they immediately and unconsciously reconstruct the unreduced form. This suggests that exemplar-based models have to assign a more prominent status to the unreduced than to the reduced exemplars of a word.

Apparently, we need a model that has characteristics of both abstractionist and exemplar-based models. Such a model assumes the storage of all tokens of a word in their broad context, but also allows generalizations over these tokens and assigns a more prominent status to unreduced than reduced tokens. This view is in line with recent studies suggesting that abstract generalizations and exemplars both affect the comprehension process (e.g., the proceedings of the special session on exemplar-based models at ICPhS 2007), possibly at different stages (Luce, McLennan & Charles-Luce 2003).

Session: Themed Panel (part 1)
Usage-based and rule based approaches to phonological variation
Friday, April 4, 2008, 13:45-15:15
room: 03