The use of sequential probabilities in the segmentation of speech

The present investigation addresses the possible utility of sequential probabilities in the segmentation of spoken language. In a series of five word- spotting and two control lexical decision experiments, high- versus low-probability consonant-vowel (Experiments 1, 2, 5, and 7) and vowel-consonant (Experiments 1, 3, 4, and 6) strings were presented either in the nonsense contexts of target words (Experiments 1-3) or within the target words themselves (Experiments 4-7). The results suggest that listeners, at least for sequences in the onset position, indeed use sequential probabilities as cues for segmentation. The probability of a sound sequence influenced segmentation more when the sequence occurred within the target words (Experiments 4-7 vs. Experiments 1-3). Furthermore, the effects were reliable only when the sequences occurred in the onset position (Experiments 1, 2, 5, and 7 vs. Experiments 1, 3, 4, and 6).     

Transitional probabilities and positional frequency phonotactics in a hierarchical model of speech segmentation

The present study explored the influence of a new metrics of phonotactics on adults’ use of transitional probabilities to segment artificial languages. We exposed French native adults to continuous streams of trisyllabic nonsense words. High-frequency words had either high or low congruence with French phonotactics, in the sense that their syllables had either high or low positional frequency in French trisyllabic words. At test, participants heard low-frequency words and part-words, which differed in their transitional probabilities (high for words, low for part-words) but were matched for frequency and phonotactic congruency. Participants’ preference for words over part-words was found only in the high-congruence languages. These results establish that subtle phonotactic manipulations can influence adults’ use of transitional probabilities to segment speech and unambiguously demonstrate that this prior knowledge interferes directly with segmentation processes, in addition to affecting subsequent lexical decisions. Implications for a hierarchical theory of segmentation cues are discussed.     

A computational model of word segmentation from continuous speech using transitional probabilities of atomic acoustic events

Word segmentation from continuous speech is a difficult task that is faced by human infants when they start to learn their native language. Several studies indicate that infants might use several different cues to solve this problem, including intonation, linguistic stress, and transitional probabilities between subsequent speech sounds. In this work, a computational model for word segmentation and learning of primitive lexical items from continuous speech is presented. The model does not utilize any a priori linguistic or phonemic knowledge such as phones, phonemes or articulatory gestures, but computes transitional probabilities between atomic acoustic events in order to detect recurring patterns in speech. Experiments with the model show that word segmentation is possible without any knowledge of linguistically relevant structures, and that the learned ungrounded word models show a relatively high selectivity towards specific words or frequently co-occurring combinations of short words.     

The impact of attention load on the use of statistical information and coarticulation as speech segmentation cues

In two artificial language learning experiments, we investigated the impact of attention load on segmenting speech through two sublexical cues: transitional probabilities (TPs) and coarticulation. In Experiment 1, we observed that coarticulation processing was resilient to high attention load, whereas TP computation was penalized in a graded manner. In Experiment 2, we showed that encouraging participants to actively search for “word” candidates enhanced overall performance but was not sufficient to preclude the impairment of statistically driven segmentation by attention load. As long as attentional resources were depleted, independently of their intention to find these “words,” participants segmented only TP words with the highest TPs, not TP words with lower TPs. Attention load thus has a graded and differential impact on the relative weighting of the cues in speech segmentation, even when only sublexical cues are available in the signal.     

The role of perceptual salience during the segmentation of connected speech

Statistical learning allows listeners to track transitional probabilities among syllable sequences and use these probabilities for subsequent speech segmentation. Recent studies have shown that other sources of information, such as rhythmic cues, can modulate the dependencies extracted via statistical computation. In this study, we explored how syllables made salient by a pitch rise affect the segmentation of trisyllabic words from an artificial speech stream by native speakers of three different languages (Spanish, English, and French). Results showed that, whereas performance of French participants did not significantly vary across stress positions (likely due to language-specific rhythmic characteristics), the segmentation performance of Spanish and English listeners was unaltered when syllables in word-initial and word-final positions were salient, but it dropped to chance level when salience was on the medial syllable. We argue that pitch rise in word-medial syllables draws attentional resources away from word boundaries, thus decreasing segmentation effectiveness.     

The monolingual nature of speech segmentation by bilinguals.

Monolingual French speakers employ a syllable-based procedure in speech segmentation; monolingual English speakers use a stress-based segmentation procedure and do not use the syllable-based procedure. In the present study French-English bilinguals participated in segmentation experiments with English and French materials. Their results as a group did not simply mimic the performance of English monolinguals with English language materials and of French monolinguals with French language materials. Instead, the bilinguals formed two groups, defined by forced choice of a dominant language. Only the French-dominant groups showed syllabic segmentation and only with French language materials. The English-dominant group showed no syllabic segmentation in either language. However, the English-dominant group showed stress-based segmentation with English language materials; the French-dominant group did not. We argue that rhythmically based segmentation procedures are mutually exclusive, as a consequence of which speech segmentation by bilinguals is, in one respect at least, functionally monolingual.     

Implicit learning of fifth- and sixth-order sequential probabilities

Serial reaction time (SRT) task studies have established that people can implicitly learn sequential contingencies as complex as fourth-order probabilities. The present study examined people’s ability to learn fifth-order (Experiment 1) and sixth-order (Experiment 2) probabilities. Remarkably, people learned fifth- and sixth-order probabilities. This suggests that the implicit sequence learning mechanism can operate over a range of at least seven sequence elements.     

Converging measures of speech segmentation in preverbal infants

Two studies using novel extensions of the conditioned head-turning method examined contributions of rhythmic and distributional properties of syllable strings to 8-month-old infants' speech segmentation. The two techniques introduced exploit fundamental, but complementary, properties of representational units. The first involved assessment of discriminative response maintenance when simple training stimuli were embedded in more complex speech contexts; the second involved measurement of infants' latencies in detecting extraneous signals superimposed on speech stimuli. A complex pattern of results is predicted if infants succeed in grouping syllables into higher-order units. Across the two studies, the predicted pattern of results emerged, indicating that rhythmic properties of speech play an important role in guiding infants toward potential linguistically relevant units and simultaneously demonstrating that the techniques proposed here provide valid, converging measures of infants' auditory representational units.     

Spontaneous segmentation in normal and in time-compressed speech

An experiment is reported in which subjects heard recorded speech passages at normal speaking rates and at various degrees of time compression, with instructions to interrupt the speech input for intermediate report by stopping the tape at points of their choosing. An analysis of subjects’ points of interruption as a function of speech rate eliminated temporal duration and number of words per segment as primary determiners of segmentation strategy. Formal linguistic constituents at all speech rates served as a good, but not perfect, predictor of segmentation boundaries. The results are interpreted in the light of on-line interactive and augmented transition network (ATN) models of speech perception.     

The active use of grammar in speech perception

Judgments of the location of short bursts of noise in sentences were used to reveal perceptual segmentation of sentences. It was assumed that segmentation would correspond to major constituent boundaries. In order to control for correlated variables of pitch and intonation, identical acoustic material was provided with alternate constituent structures. It was found that differences in response to identical strings were predicted by the points of variation in constituent structure.     

How to use probabilities in reasoning

Conclusions Probabilities are important in belief updating, but probabilistic reasoning does not subsume everything else (as the Bayesian would have it). On the contrary, Bayesian reasoning presupposes knowledge that cannot itself be obtained by Bayesian reasoning, making generic Bayesianism an incoherent theory of belief updating. Instead, it is indefinite probabilities that are of principal importance in belief updating. Knowledge of such indefinite probabilities is obtained by some form of statistical induction, and inferences to non-probabilistic conclusions are carried out in accordance with the statistical syllogism. Such inferences have been the focus of much attention in the nonmonotonic reasoning literature, but the logical complexity of such inference has not been adequately appreciated.     

The use of rhythm in attending to speech

Three experiments examined attentional allocation during speech processing to determine whether listeners capitalize on the rhythmic nature of speech and attend more closely to stressed than to unstressed syllables. Ss performed a phoneme monitoring task in which the target phoneme occurred on a syllable that was either predicted to be stressed or unstressed by the context preceding the target word. Stimuli were digitally edited to eliminate the local acoustic correlates of stress. A sentential context and a context composed of word lists, in which all the words had the same stress pattern, were used. In both cases, the results suggest that attention may be preferentially allocated to stressed syllables during speech processing. However, a normal sentence context may not provide strong predictive cues to lexical stress, limiting the use of the attentional focus.     

Effects of syntactic expectations on speech segmentation

Although the effect of acoustic cues on speech segmentation has been extensively investigated, the role of higher order information (e.g., syntax) has received less attention. Here, the authors examined whether syntactic expectations based on subject-verb agreement have an effect on segmentation and whether they do so despite conflicting acoustic cues. Although participants detected target words faster in phrases containing adequate acoustic cues ("spins" in take spins and "pins" in takes pins), this acoustic effect was suppressed when the phrases were appended to a plural context (those women take spins/*takes pins [with the asterisk indicating a syntactically unacceptable parse]). The syntactically congruent target ("spins") was detected faster regardless of the acoustics. However, a singular context (that woman *take spins/takes pins) had no effect on segmentation, and the results resembled those of the neutral phrases. Subsequent experiments showed that the discrepancy was due to the relative time course of syntactic expectations and acoustics cues. Taken together, the data suggest that syntactic knowledge can facilitate segmentation but that its effect is substantially attenuated if conflicting acoustic cues are encountered before full realization of the syntactic constraint.     

Possible words and fixed stress in the segmentation of Slovak speech

The possible-word constraint (PWC; Norris, McQueen, Cutler, & Butterfield, 1997) has been proposed as a language-universal segmentation principle: Lexical candidates are disfavoured if the resulting segmentation of continuous speech leads to vowelless residues in the input—for example, single consonants. Three word-spotting experiments investigated segmentation in Slovak, a language with single-consonant words and fixed stress. In Experiment 1, Slovak listeners detected real words such as ruka “hand” embedded in prepositional-consonant contexts (e.g., /gruka/) faster than those in nonprepositional-consonant contexts (e.g., /truka/) and slowest in syllable contexts (e.g., /dugruka/). The second experiment controlled for effects of stress. Responses were still fastest in prepositional-consonant contexts, but were now slowest in nonprepositional-consonant contexts. In Experiment 3, the lexical and syllabic status of the contexts was manipulated. Responses were again slowest in nonprepositional-consonant contexts but equally fast in prepositional-consonant, prepositional-vowel, and nonprepositional-vowel contexts. These results suggest that Slovak listeners use fixed stress and the PWC to segment speech, but that single consonants that can be words have a special status in Slovak segmentation. Knowledge about what constitutes a phonologically acceptable word in a given language therefore determines whether vowelless stretches of speech are or are not treated as acceptable parts of the lexical parse.     

Overnight lexical consolidation revealed by speech segmentation

Two experiments explored the consolidation of spoken words, and assessed whether post-sleep novel competitor effects truly reflect engagement of these novel words in competition for lexical segmentation. Two types of competitor relationships were contrasted: the onset-aligned case (such as "frenzylk"), where the novel word is a close variant of the existing word: they start at the same time point and overlap on most of their segments; and the fully embedding case (such as "lirmucktoze"), where the existing word corresponds to a smaller embedded portion of its novel competitor and is thus less noticeable. Experiment 1 (pause detection) revealed a similar performance for both cases, with no competitor effect immediately after exposure, but significant inhibition after 24 h and seven days. Experiment 2 (word spotting) produced exactly the same pattern; however, as is the case with existing word carriers (cf. McQueen, Norris, & Cutler, 1994), the inhibition was much stronger for fully embedded than for onset-aligned targets (e.g., "lirmuckt" vs. "frenzyl"). Meanwhile, explicit measures of learning, i.e., free recall and recognition, improved over time. These results cannot be explained by either consolidation of episodic traces or acquisition of new phonological/dialectal variants. We argue instead that they reflect a general trait of vocabulary learning and consolidation.     

An interaction between prosody and statistics in the segmentation of fluent speech

Sensitivity to prosodic cues might be used to constrain lexical search. Indeed, the prosodic organization of speech is such that words are invariably aligned with phrasal prosodic edges, providing a cue to segmentation. In this paper we devise an experimental paradigm that allows us to investigate the interaction between statistical and prosodic cues to extract words from a speech stream. We provide evidence that statistics over the syllables are computed independently of prosody. However, we also show that trisyllabic sequences with high transition probabilities that straddle two prosodic constituents appear not to be recognized. Taken together, our findings suggest that prosody acts as a filter, suppressing possible word-like sequences that span prosodic constituents.     

Lexical viability constraints on speech segmentation by infants

The Possible Word Constraint limits the number of lexical candidates considered in speech recognition by stipulating that input should be parsed into a string of lexically viable chunks. For instance, an isolated single consonant is not a feasible word candidate. Any segmentation containing such a chunk is disfavored. Five experiments using the head-turn preference procedure investigated whether, like adults, 12-month-olds observe this constraint in word recognition. In Experiments 1 and 2, infants were familiarized with target words (e.g., rush), then tested on lists of nonsense items containing these words in "possible" (e.g., "niprush" [nip+rush]) or "impossible" positions (e.g., "prush" [p+rush]). The infants listened significantly longer to targets in "possible" versus "impossible" contexts when targets occurred at the end of nonsense items (rush in "prush"), but not when they occurred at the beginning (tan in "tance"). In Experiments 3 and 4, 12-month-olds were similarly familiarized with target words, but test items were real words in sentential contexts (win in "wind" versus "window"). The infants listened significantly longer to words in the "possible" condition regardless of target location. Experiment 5 with targets at the beginning of isolated real words (e.g., win in "wind") replicated Experiment 2 in showing no evidence of viability effects in beginning position. Taken together, the findings suggest that, in situations in which 12-month-olds are required to rely on their word segmentation abilities, they give evidence of observing lexical viability constraints in the way that they parse fluent speech.     

The use of conditional probabilities in applications of Holland's theory

The hexagon of J. L. Holland (1973, Making Vocational Choices: A Theory of Careers, Englewood Cliffs, NJ, Prentice-Hall) is a useful heuristic device for predicting congruency and consistency among the six personality/environment types specified by his theory. One limitation of the traditional hexagon is the implicit bidirectionality: the hexagon displays a single distance between two types. People in applied areas may assume a symmetry which does not exist. Examples are given of situations in which a consistent asymmetry occurs; e.g., in the illustrative data presented, C (Conventional) types are more likely to have E (Enterprising) as a secondary interest than E types are to have C as a secondary interest. Implications of such asymmetries are discussed and suggestions made concerning counseling and other applications.     

Bootstrapping the lexicon: a computational model of infant speech segmentation

Prelinguistic infants must find a way to isolate meaningful chunks from the continuous streams of speech that they hear. BootLex, a new model which uses distributional cues to build a lexicon, demonstrates how much can be accomplished using this single source of information. This conceptually simple probabilistic algorithm achieves significant segmentation results on various kinds of language corpora - English, Japanese, and Spanish; child- and adult-directed speech, and written texts; and several variations in coding structure - and reveals which statistical characteristics of the input have an influence on segmentation performance. BootLex is then compared, quantitatively and qualitatively, with three other groups of computational models of the same infant segmentation process, paying particular attention to functional characteristics of the models and their similarity to human cognition. Commonalities and contrasts among the models are discussed, as well as their implications both for theories of the cognitive problem of segmentation itself, and for the general enterprise of computational cognitive modeling.