Proceedings of 3rd European conference on speech communication and technology, Eurospeech '93, Berlin, Germany, 21-23 September 1993. ARE STRESS AND PHONEMIC STRING PROCESSED SEPARATELY? EVIDENCE FROM SPEECH ILLUSIONS Valérie Pasdeloup, José Morais and Régine Kolinsky Laboratoire de Psychologie Expérimentale, Université Libre de Bruxelles, Brussels, Belgium ABSTRACT Tins paper describes a perceptual experiment which was run to determine if lexical stress pattern and phonemic string are represented separately at some level of processing. We used a target word recognition under dichotic listening presentation: two stimuli are presented simultaneously, one to one ear and one to the other ear. The stimuli are constructed in such a way that, if listeners combine the phonemic string from the stimulus presented to one ear with the stress pattern from the stimulus presented to the other ear, they experience the illusory perception of a word which corresponds to the prespecified target. The results suggest that stress pattern and segmental string can be represented separately at least at a prelexical level. Keywords : lexical stress, prosody, speech processing. 1. INTRODUCTION One important issue in cognitive psycholinguistic is the psychological reality of the entities postulated by linguists and phonologists. Much work is aimed at discovering how the acoustic speech stream is mapped onto mental representations of words (the so-called mental lexicon). Most of the models of spoken word recognition assume that tlie speech stream is initially analyzed as a sequence of prelexical units such as phonemes and syllables. Besides prelexical units, other elements relative to more global phonological domains might play a role in word recognition, One of them is stress pattern. In lexical-stress languages such as English and Spanish, the placement of stress is not fixed, and lexical stress patterns can convey information about word identity. Research based on production and perception errors (slips of the ear and of the tongue, and tips- of-the-tongue) suggests that lexical stress is part of word's phonological representation as much as its segmental constituents III. Models of spoken word recognition do not generally consider global information such as the lexical stress pattern or other prosodie structures as a potential source of information. However, there is clear evidence, namely from experiments on mis-stressing, that lexical stress influences word recognition. For instance, Cutler and Clifton 121 found that mis-stressing hampers bisyllabic word recognition, even when the stress contrast does not involve a vowel quality difference. However, results on the effect of mis-stressing can be opposed to those on the effect of prior information about stress on lexical access. Both Cutler and Clifton 111 and Cutler fil found that prior knowledge of lexical stress does not facilitate lexical access. In a cross-modal priming task, pairs of words which have the same phonemic constituents but a different stress pattern, like FORbear- forBEAR, produce the priming effects characteristics of homophones /3/ (throughout this article, upper case will be used to represent a stressed syllable). The lack of coherence of the results across tasks may be only apparent. Different experimental situations may yield different outcomes if they tap different stages of processing and if lexical stress is only influential in some of these. Lexical stress information could be used at least during two stages of word recognition. At the prelexical stage of processing, stress could constrain lexical access so that a word stimulus having a given stress pattern would not contact a word representation having the same phonemic constituents but a different stress pattern (FORbear would not access forBEAR). At the postiexical stage, stress could be used when lexical access has already been achieved in order to perform a definite lexical choice. The fact that FORbear does activate forBEAR and vice-versa, strongly suggests that lexical stress influences word recognition only at the postlexical stage. Although the stress pattern probably influences word recognition at a relatively late stage, it may have been extracted much earlier. A further issue consitutes the topic of the present work: Are the phonemic string and the lexical stress pattern extracted separately! Can tlie stress pattern be represented, at some time, in isolation, regardless of the identity of the phonemic string, or, on the contrary, is its representation constantly associated to the representation of the phonemic string? In nonlinear phonological theories (metrical and autosegmental), suprasegmentals such as lexical stress pattern and tone pattern are independent of tlie phonemic strings to which they are associated 141. In the present experiment, we investigated whether in Portuguese, a lexical-stress language, the stress pattern and tlie phonemic string are represented separately at some processing level. We used a target detection task that was originally designed for the study of the units of spoken word recognition, such as voicing and place features 151. It uses the dichotic listening presentation, where two different stimuli are presented simultaneously, one to one ear and one to the other ear. The stimuli are constructed in such a way that, if