Models Of Spoken Word Recognition Research Articles

Integrating Form and Meaning: A Distributed Model of Speech Perception

We present a new distributed connectionist model of the perception of spoken words. The model employs a representation of speech that combines lexical information with abstract phonological information, with lexical access modelled as a direct mapping onto this single distributed representation. We first examine the integration of partial cues to phonological identity, showing that the model provides a sound basis for simulating phonetic and lexical decision data from Marslen-Wilson and Warren (1994). We then investigate the time course of lexical access, and argue that the process of competition between word candidates during lexical access can be interpreted in terms of interference between distributed lexical representations. The relation between our model and other models of spoken word recognition is discussed.

Word recognition from acoustic onsets and acoustic offsets: Effects of cohort size and syllabic stress

ABSTRACTIn the traditional gating technique, subjects hear increasing amounts of word-onset information from spoken words until the words can be correctly identified. The experiment reported here contrasted word-onset gating with results when words were gated from their word endings. A significant recognition advantage for words gated from their onsets was demonstrated. This effect was eliminated, however, when we took into account the number of word possibilities that shared overlapping phonology and the same stress pattern as the target words at their recognition points. These results support the position that the perceptual advantage of word-initial information can be understood within a general goodness-of-fit model of spoken word recognition.

Representational specificity of lexical form: Implications for models of spoken word recognition

Much recent work on spoken-word recognition has suggested that the representation of phonological form in memory is quite specific and veridical, coding such lexically nondistinctive information as talker voice and allophonic variation. The implications of this research for current models of spoken word recognition will be discussed. In particular, exemplar-based approaches to modeling spoken word recognition and the need for multiple levels of representation of sound patterns in memory, ranging from specific, instance-based representations to more general, abstract phonological forms will also be discussed.

Building lexical neighborhoods.

The Neighborhood Activation Model (NAM) states that the speed and accuracy of spoken word recognition are a function of the number and nature of neighbors—or similar sounding words—activated in memory by stimulus input. In particular, the model states that spoken word recognition is a function of (1) target word frequency, (2) similarity neighborhood density, and (3) similarity neighborhood frequency. In a pretest–training –posttest design, these three factors were manipulated using nonword target words and nonword neighbors to test directly predictions of NAM. The results were only partially consistent with the model. As predicted, processing times to high‐frequency targets increased as a function of neighborhood density. However, processing times for low‐frequency targets actually decreased with increased neighborhood density. Furthermore, no significant effects of neighborhood frequency were obtained. The implications of these findings for NAM and other models of spoken word recognition are discussed.

The structural organization of the mental lexicon and its contribution to age-related declines in spoken word recognition

Three experiments were conducted within the framework of the Neighborhood Activation Model of spoken word recognition to examine how the structural organization of the mental lexicon may contribute to age-related declines in spoken language processing. According to the model, lexically hard words are items that are phonetically similar to many other high-frequency words and should be identified less accurately than easy words which are phonetically similar to relatively few, low-frequency items. The results of experiment 1 demonstrated that lexical difficulty had differential effects on older and younger listeners. Specifically, relative to young listeners, older adults exhibited significantly poorer identification scores for lexically hard items. Experiment 2 indicated that these age differences in the effects of lexical difficulty were maintained under conditions in which performance for easy words was the same for older and younger listeners. In the third experiment, reducing the resources available for perceptual identification, by changing from single to multiple talkers, produced greater effects of lexical difficulty for older, compared to younger, listeners. Explanations for the results based on diminished cognitive resources and impaired inhibitory control will be discussed. [Work supported by Washington Univ.]

Effects of sensory information and processing time in spoken-word recognition

Abstract Current models of spoken-word recognition describe access to lexical representations in terms of activation and decay. This research investigated an important aspect of activation, that is the impact of processing time. In the natural situation of continuous speech, the more acoustic information a listener receives about a word, the more time is available to process earlier parts of this input. These two naturally confounded aspects of spoken-word processing—sensory input and time—were investigated in a cross-modal priming study. The results showed a separable impact of time and signal on the activational state of lexical elements.

Getting it right? Using aphasic naming errors to evaluate theoretical models of spoken word recognition

Abstract Different models of spoken word production make different predictions regarding the extent of effects of certain word properties on the output of that model. These predictions are examined with regard to the effect of these variables on the production of semantic and phonological errors by aphasic subjects. Thus the production of semantic errors is found to be affected by imageability and the production of phonological errors by word length, but not vice versa. It is argued that this pattern of variables affecting the production of semantic and phonological errors is better explained by models which require strictly sequential stage-by-stage processing (e.g. Levelt et al., 1991a; Morton, 1970; 1979) than by multi-layer perceptron (e.g. Plaut & Shallice, 1991; 1993) or interactive activation models (e.g. Dell, 1986; 1989).

Voice information in auditory form-based priming

In auditory form-based priming, subjects are presented with speech stimuli that are phonetically related. Research has shown that a prime can facilitate recognition of a target item if prime and target share an initial phonetic segment. This facilitory effect has been considered either a result of residual activation of the initial phonetic representation [L. M. Slowiaczek and M. B. Hamburger, JEP: LMC 18, 1239–1250 (1992)] or post-perceptual guessing strategy [Goldinger etal., JEP: LMC 18, 1211–1238 (1992)]. It could be, however, that the facilitation is at least in part based on nonlinguistic auditory information. Previous studies have always used the same speaker for prime and target items. In the present investigation, auditory form-based priming experiments were conducted where voice information was changed from prime to target. Initial results reveal that varying voice information eliminates the facilitation effect. This could suggest that the effect is based on residual auditory activation. Follow-up experiments test whether eliminating the acoustic trace through backward masking attenuates the facilitation effect. Implications of these results will be discussed in terms of current models of spoken word recognition. [Work supported by NSF.]

Rhyme decisions to spoken words and nonwords.

Lexical effects in auditory rhyme-decision performance were examined in three experiments. Experiment 1 showed reliable lexical involvement: rhyme-monitoring responses to words were faster than rhyme-monitoring responses to nonwords; and decisions were faster in response to high-frequency as opposed to low-frequency words. Experiments 2 and 3 tested for lexical influences in the rejection of three types of nonrhyming item: words, nonwords with rhyming lexical neighbor (e.G., jop after the cue rob), and nonwords with no rhyming lexical neighbor (e.g., vop after rob). Words were rejected more rapidly than nonwords, and there were reliable differences in the speed and accuracy of rejection of the two types of nonword. The advantage for words over nonwords was replicated for positive rhyme decisions. However, there were no differences in the speed of acceptance, as rhymes, of the two types of nonword. The implications of these results for interactive and autonomous models of spoken word recognition are discussed. It is concluded that the differences in rejection of nonrhyming nonwords are due to the operation of a guessing strategy.

Hemispheric differences in spoken word recognition

Previous research on visual word recognition has demonstrated differences in the manner that the left and right hemispheres process lexically ambiguous words [Burgess and Simpson, Brain Language 33, 86–103 (1988)]. The present study evaluated these findings for spoken word recognition. Lexically ambiguous primes (e.g., BANK) were presented binaurally. A target word was then presented monaurally to the left or right ear for a speeded lexical decision response (i.e., WORD-NON-WORD). Target words were either (1) related to the dominant meaning of the ambiguous prime (BANK-MONEY), (2) related to the subordinate meaning of the prime (BANK-RIVER), or (3) unrelated to the prime. Priming of dominant meanings produced more facilitation than priming of subordinate meanings regardless of ear of presentation. In addition, a right-ear advantage was observed for words but not for non-words. Implications of these results for models of spoken word recognition will be discussed.

Priming lexical neighbors of spoken words: Effects of competition and inhibition

Two experiments employing an auditory priming paradigm were conducted to test predictions of the Neighborhood Activation Model of spoken word recognition ( Luce & Pisoni, 1989, Neighborhoods of words in the mental lexicon. Manuscript under review). Acousticphonetic similarity, neighborhood densities, and frequencies of prime and target words were manipulated. In Experiment 1, priming with low frequency, phonetically related spoken words inhibited target recognition, as predicted by the Neighborhood Activation Model. In Experiment 2, the same prime-target pairs were presented with a longer inter-stimulus interval and the effects of priming were eliminated. In both experiments, predictions derived from the Neighborhood Activation Model regarding the effects of neighborhood density and word frequency were supported. The results are discussed in terms of competing activation of lexical neighbors and the dissociation of activation and frequency in spoken word recognition.

Accessing spoken words: The importance of word onsets.

Approaches to spoken word recognition differ in the importance they assign to word onsets during lexical access. This research contrasted the hypothesis that lexical access is strongly directional with the hypothesis that word onsets are less important than the overall goodness of fit between input and lexical form. A cross-modal priming technique was used to investigate the extent to which a rhyme prime (a prime that differs only in its first segment from the word that is semantically associated with the visual probe) is as effective a prime as the original word itself. Earlier research had shown that partial primes that matched from word onset were very effective cross-modal primes. The present results show that, irrespective of whether the rhyme prime was a real word or not, and irrespective of the amount of overlap between the rhyme prime and the original word, the rhymes are much less effective primes than the full word. In fact, no overall priming effect could be detected at all except under conditions in which the competitor environment was very sparse. This suggests that word onsets do have a special status in the lexical access of spoken words. A fundamental property of the speech signal is its intrinsic directionality in time. Spoken utterances are spread out along the time line, moving necessarily from beginning to end, in a way that is not true of written language. This directionality of the speech input is strongly reflected in the claims made by the cohort model of spoken word recognition for the manner in which speech inputs are mapped onto the representations of word forms in the mental lexicon (Marslen-Wilson, 1984, 1987; Marslen-Wilson & Tyler, 1980; Marslen-Wilson & Welsh, 1978; Tyler, 1984; Tyler & Wessels, 1983; Warren & Marslen-Wilson, 1987). The cohort model of word recognition stresses the sequential and continuous nature of the mapping between the speech input and mental representations of word forms. This emphasis is closely tied up with the concept of a cohort and its implications for the properties of the on-line lexical decision space. In particular, according to the cohort model, the decision space is determined by the beginnings of words. The speech input at the beginning of the word maps onto all lexical items that share the same initial sequence. This initial set of candidates is termed the word-initial cohort, and the subsequent process of word recognition is determined by the

Malleable frequency effects in spoken word recognition

Recent research has demonstrated that word frequency effects in spoken word recognition are primarily due to processes responsible for deciding among multiple lexical items activated in memory and not to differences in activation levels or thresholds of word units. The present study further supports this claim by demonstrating that frequency effects in the identification of spoken words are malleable and can be modified by varying the composition of the experimental stimuli. A set of high-frequeocy words was presented for identification mixed with either other high-frequency words or with low-frequency words. A set of low-frequency words was also presented with high- or low-frequency words. The results demonstrated that high-frequency words were identified more accurately when presented with other high-frequency words and that low-frequency words were identified more accurately when presented with other low-frequency words. The implications of these results for current models of spoken word recognition, in particular the neighborhood activation model, will be discussed. [Work supported by NIH Grant NS-12179 to Indiana University.]

Gaining insight into the representational unit for spoken words with a naming task

Theoretical proposals for the representational unit used in recognizing spoken words include the phoneme, allophone, and position-specific phoneme. The present experiment was devised to differentiate among these proposals. Subjects heard natural CVC stimuli in prime-target pairs and named the second item (target) as quickly as possible. The phonetic overlap between prime and target was varied to create five different intrapair relations such that the expected pattern of priming effects across the set of relations was different for each of the three phonetic representations. The obtained pattern is inconsistent with an abstract phonemic unit and points to either a position-specific phonemic or allophonic unit. This pattern was found for both word and nonword stimuli. Thus these priming results appear to tap the recognition process prior to lexical activation. Priming effects were eliminated when stimuli within a prime-target pair were produced by different voices. The implications of the findings for refining models of spoken word recognition will be discussed. [Work supported by NINCDS.]

Continuous uptake of acoustic cues in spoken word recognition.

The cohort model of spoken word recognition postulates a maximally efficient uptake of acousticphonetic information during lexical access and selection. To test this, we used the gating paradigm in two experiments to measure the effects of anticipatory vowel coarticulation in word-final vowel-consonant transitions on the recognition of spoken words. In Experiment 1, we tested place contrasts between plosives and manner contrasts among plosives, nasals, and fricatives; we found a continuous effect of coarticulatory changes in vowel spectra on the process of lexical choice. In Experiment 2, we included word frequency as a variable and found that the effects of frequency operate independently of coarticulatory effects. These results are inconsistent with views of spoken word recognition that insist on discrete phoneme- or syllable-sized delays in the projection of acoustic-phonetic information onto form representations in the mental lexicon.

Prosodic structure and spoken word recognition

The aim of this paper is to call attention to the role played by prosodic structure in continuous word recognition. First we argue that the written language notion of the word has had too much impact on models of spoken word recognition. Next we discuss various characteristics of prosodic structure that bear on processing issues. Then we present a view of continuous word recognition which takes into account the alternating pattern of weak and strong syllables in the speech stream. A lexical search is conducted with the stressed syllables while the weak syllables are identified through a pattern-recognition-like analysis and the use of phonotactic and morphonemic rules. We end by discussing the content word vs. function word access controversy in the light of our view.

Visual and Auditory Recognition of Prefixed Words

The involvement of stem storage and prefix stripping in the recognition of spoken and printed prefixed words was examined. In both an auditory and a visual lexical decision experiment, it was found that prefixed nonwords were more difficult to classify as nonwords than were non-prefixed nonwords. This difference was larger, though, when the “stem” of the nonword was a genuine stem in English (e.g., dejoice versus tejoice) than when it was not (e.g., dejouse versus tejouse). The results suggest that prefixed words are recognized via a representation of their stem after the prefix has been removed, and this is true regardless of the modality of presentation of the word. Implications are considered for the Cohort model of spoken word recognition.

Exploring the cohort model of spoken word recognition

Predictions derived from the Cohort Model of spoken word recognition were tested in four experiments using an auditory lexical decision task. The first experiment produced results that were compatible with the model, in that the point at which a word could be uniquely identified appeared to influence reaction times. The second and third experiments demonstrated that the processing of a nonword phoneme string continues after the point at which there are no possible continuations that would make a word. The number of phonemes following the point of deviation from a word was shown to affect reaction times, as well as the similarity of the nonword to a word. The final experiment demonstrated a frequency effect when high and low frequency words were matched on their point of unique identity. These last three results are not consistent with the Cohort Model and so an alternative account is put forward. According to this account, the first few phonemes are used to activate all words beginning with those phonemes and then these candidates are checked back to the original stimulus. This model provides greater flexibility than the Cohort Model and allows for mispronounced and misperceived words to be correctly recognized.

Cross-linguistic approaches to lexical segmentation

This paper examines the problem of lexical segmentation in order to identify language-specific and language-independent properties of lexical processing systems. A potential universal of lexical segmentation is first evaluated for its efficacy across languages. This postlexical segmentation strategy assumes that listeners recognize each word rapidly enough to be able to predict both its offset and the onset of the following word. It is argued that this strategy cannot account for every lexical segmentation decision in utterances in any language. Listeners must also use other types of segmentation information. Two types of segmentation information, distributional and relational, which appear to vary crosslinguistically, are presented with examples. To exploit this information, listeners must infer word boundaries or boundaries of other domains from events or sequences of events in the signal. The problem posed for current models of spoken word recognition by the integration of such segmentation information with that from the putative postlexical strategy is discussed. © 1985, Mouton Publishers.