Speech segmentation by statistical learning is supported by domain-general processes within working memory

The purpose of this study was to examine the extent to which working memory resources are recruited during statistical learning (SL). Participants were asked to identify novel words in an artificial speech stream where the transitional probabilities between syllables provided the only segmentation cue. Experiments 1 and 2 demonstrated that segmentation performance improved when the speech rate was slowed down, suggesting that SL is supported by some form of active processing or maintenance mechanism that operates more effectively under slower presentation rates. In Experiment 3 we investigated the nature of this mechanism by asking participants to perform a two-back task while listening to the speech stream. Half of the participants performed a two-back rhyme task designed to engage phonological processing, whereas the other half performed a comparable two-back task on un-nameable visual shapes. It was hypothesized that if SL is dependent only upon domain-specific processes (i.e., phonological rehearsal), the rhyme task should impair speech segmentation performance more than the shape task. However, the two loads were equally disruptive to learning, as they both eradicated the benefit provided by the slow rate. These results suggest that SL is supported by working-memory processes that rely on domain-general resources.     

Speech segmentation by statistical learning depends on attention

We addressed the hypothesis that word segmentation based on statistical regularities occurs without the need of attention. Participants were presented with a stream of artificial speech in which the only cue to extract the words was the presence of statistical regularities between syllables. Half of the participants were asked to passively listen to the speech stream, while the other half were asked to perform a concurrent task. In Experiment 1, the concurrent task was performed on a separate auditory stream (noises), in Experiment 2 it was performed on a visual stream (pictures), and in Experiment 3 it was performed on pitch changes in the speech stream itself. Invariably, passive listening to the speech stream led to successful word extraction (as measured by a recognition test presented after the exposure phase), whereas diverted attention led to a dramatic impairment in word segmentation performance. These findings demonstrate that when attentional resources are depleted, word segmentation based on statistical regularities is seriously compromised.     

The link between statistical segmentation and word learning in adults

Many studies have shown that listeners can segment words from running speech based on conditional probabilities of syllable transitions, suggesting that this statistical learning could be a foundational component of language learning. However, few studies have shown a direct link between statistical segmentation and word learning. We examined this possible link in adults by following a statistical segmentation exposure phase with an artificial lexicon learning phase. Participants were able to learn all novel object-label pairings, but pairings were learned faster when labels contained high probability (word-like) or non-occurring syllable transitions from the statistical segmentation phase than when they contained low probability (boundary-straddling) syllable transitions. This suggests that, for adults, labels inconsistent with expectations based on statistical learning are harder to learn than consistent or neutral labels. In contrast, a previous study found that infants learn consistent labels, but not inconsistent or neutral labels.     

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.     

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.     

Social cues modulate the representations underlying cross-situational learning

Because children hear language in environments that contain many things to talk about, learning the meaning of even the simplest word requires making inferences under uncertainty. A cross-situational statistical learner can aggregate across naming events to form stable word-referent mappings, but this approach neglects an important source of information that can reduce referential uncertainty: social cues from speakers (e.g., eye gaze). In four large-scale experiments with adults, we tested the effects of varying referential uncertainty in cross-situational word learning using social cues. Social cues shifted learners away from tracking multiple hypotheses and towards storing only a single hypothesis (Experiments 1 and 2). In addition, learners were sensitive to graded changes in the strength of a social cue, and when it became less reliable, they were more likely to store multiple hypotheses (Experiment 3). Finally, learners stored fewer word-referent mappings in the presence of a social cue even when given the opportunity to visually inspect the objects for the same amount of time (Experiment 4). Taken together, our data suggest that the representations underlying cross-situational word learning of concrete object labels are quite flexible: In conditions of greater uncertainty, learners store a broader range of information.     

Newborns are sensitive to multiple cues for word segmentation in continuous speech

Before infants can learn words, they must identify those words in continuous speech. Yet, the speech signal lacks obvious boundary markers, which poses a potential problem for language acquisition (Swingley, Philos Trans R Soc Lond. Series B, Biol Sci 364 (1536), 3617–3632, 2009). By the middle of the first year, infants seem to have solved this problem (Bergelson & Swingley, Proc Natl Acad Sci 109 (9), 3253–3258, 2012; Jusczyk & Aslin, Cogn Psychol 29 , 1–23, 1995), but it is unknown if segmentation abilities are present from birth, or if they only emerge after sufficient language exposure and/or brain maturation. Here, in two independent experiments, we looked at two cues known to be crucial for the segmentation of human speech: the computation of statistical co‐occurrences between syllables and the use of the language's prosody. After a brief familiarization of about 3 min with continuous speech, using functional near‐infrared spectroscopy, neonates showed differential brain responses on a recognition test to words that violated either the statistical (Experiment 1) or prosodic (Experiment 2) boundaries of the familiarization, compared to words that conformed to those boundaries. Importantly, word recognition in Experiment 2 occurred even in the absence of prosodic information at test, meaning that newborns encoded the phonological content independently of its prosody. These data indicate that humans are born with operational language processing and memory capacities and can use at least two types of cues to segment otherwise continuous speech, a key first step in language acquisition.     

Timing is everything: Changes in presentation rate have opposite effects on auditory and visual implicit statistical learning

Implicit statistical learning (ISL) is exclusive to neither a particular sensory modality nor a single domain of processing. Even so, differences in perceptual processing may substantially affect learning across modalities. In three experiments, statistically equivalent auditory and visual familiarizations were presented under different timing conditions that either facilitated or disrupted temporal processing (fast or slow presentation rates). We find an interaction of rate and modality of presentation: At fast rates, auditory ISL was superior to visual. However, at slow presentation rates, the opposite pattern of results was found: Visual ISL was superior to auditory. Thus, we find that changes to presentation rate differentially affect ISL across sensory modalities. Additional experiments confirmed that this modality-specific effect was not due to cross-modal interference or attentional manipulations. These findings suggest that ISL is rooted in modality-specific, perceptually based processes.     

The metamorphosis of the statistical segmentation output: Lexicalization during artificial language learning

This study combined artificial language learning (ALL) with conventional experimental techniques to test whether statistical speech segmentation outputs are integrated into adult listeners’ mental lexicon. Lexicalization was assessed through inhibitory effects of novel neighbors (created by the parsing process) on auditory lexical decisions to real words. Both immediately after familiarization and post-one week, ALL outputs were lexicalized only when the cues available during familiarization (transitional probabilities and wordlikeness) suggested the same parsing (Experiments 1 and 3). No lexicalization effect occurred with incongruent cues (Experiments 2 and 4). Yet, ALL differed from chance, suggesting a dissociation between item knowledge and lexicalization. Similarly contrasted results were found when frequency of occurrence of the stimuli was equated during familiarization (Experiments 3 and 4). Our findings thus indicate that ALL outputs may be lexicalized as far as the segmentation cues are congruent, and that this process cannot be accounted for by raw frequency.     

Learning diphone-based segmentation

This paper reconsiders the diphone-based word segmentation model of Cairns, Shillcock, Chater, and Levy (1997) and Hockema (2006), previously thought to be unlearnable. A statistically principled learning model is developed using Bayes' theorem and reasonable assumptions about infants' implicit knowledge. The ability to recover phrase-medial word boundaries is tested using phonetic corpora derived from spontaneous interactions with children and adults. The (unsupervised and semi-supervised) learning models are shown to exhibit several crucial properties. First, only a small amount of language exposure is required to achieve the model's ceiling performance, equivalent to between 1 day and 1 month of caregiver input. Second, the models are robust to variation, both in the free parameter and the input representation. Finally, both the learning and baseline models exhibit undersegmentation, argued to have significant ramifications for speech processing as a whole.     

Linking sounds to meanings: Infant statistical learning in a natural language

The processes of infant word segmentation and infant word learning have largely been studied separately. However, the ease with which potential word forms are segmented from fluent speech seems likely to influence subsequent mappings between words and their referents. To explore this process, we tested the link between the statistical coherence of sequences presented in fluent speech and infants’ subsequent use of those sequences as labels for novel objects. Notably, the materials were drawn from a natural language unfamiliar to the infants (Italian). The results of three experiments suggest that there is a close relationship between the statistics of the speech stream and subsequent mapping of labels to referents. Mapping was facilitated when the labels contained high transitional probabilities in the forward and/or backward direction (Experiment 1). When no transitional probability information was available (Experiment 2), or when the internal transitional probabilities of the labels were low in both directions (Experiment 3), infants failed to link the labels to their referents. Word learning appears to be strongly influenced by infants’ prior experience with the distribution of sounds that make up words in natural languages.     

Infants learn phonotactic regularities from brief auditory experience

Two experiments investigated whether novel phonotactic regularities, not present in English, could be acquired by 16.5-month-old infants from brief auditory experience. Subjects listened to consonant-vowel-consonant syllables in which particular consonants were artificially restricted to either initial or final position (e.g. /baep/ not /paeb/). In a later head-turn preference test, infants listened longer to new syllables that violated the experimental phonotactic constraints than to new syllables that honored them. Thus, infants rapidly learned phonotactic regularities from brief auditory experience and extended them to unstudied syllables, documenting the sensitivity of the infant's language processing system to abstractions over linguistic experience.     

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.     

Cue integration with categories: Weighting acoustic cues in speech using unsupervised learning and distributional statistics

During speech perception, listeners make judgments about the phonological category of sounds by taking advantage of multiple acoustic cues for each phonological contrast. Perceptual experiments have shown that listeners weight these cues differently. How do listeners weight and combine acoustic cues to arrive at an overall estimate of the category for a speech sound? Here, we present several simulations using a mixture of Gaussians models that learn cue weights and combine cues on the basis of their distributional statistics. We show that a cue‐weighting metric in which cues receive weight as a function of their reliability at distinguishing phonological categories provides a good fit to the perceptual data obtained from human listeners, but only when these weights emerge through the dynamics of learning. These results suggest that cue weights can be readily extracted from the speech signal through unsupervised learning processes.     

Stress changes the representational landscape: evidence from word segmentation

Over the past couple of decades, research has established that infants are sensitive to the predominant stress pattern of their native language. However, the degree to which the stress pattern shapes infants' language development has yet to be fully determined. Whether stress is merely a cue to help organize the patterns of speech or whether it is an important part of the representation of speech sound sequences has still to be explored. Building on research in the areas of infant speech perception and segmentation, we asked how several months of exposure to the target language shapes infants' speech processing biases with respect to lexical stress. We hypothesize that infants represent stressed and unstressed syllables differently, and employed analyses of child-directed speech to show how this change to the representational landscape results in better distribution-based word segmentation as well as an advantage for stress-initial syllable sequences. A series of experiments then tested 9- and 7-month-old infants on their ability to use lexical stress without any other cues present to parse sequences from an artificial language. We found that infants adopted a stress-initial syllable strategy and that they appear to encode stress information as part of their proto-lexical representations. Together, the results of these studies suggest that stress information in the ambient language not only shapes how statistics are calculated over the speech input, but that it is also encoded in the representations of parsed speech sequences.     

Effects of visual information on adults' and infants' auditory statistical learning

Infant and adult learners are able to identify word boundaries in fluent speech using statistical information. Similarly, learners are able to use statistical information to identify word-object associations. Successful language learning requires both feats. In this series of experiments, we presented adults and infants with audio-visual input from which it was possible to identify both word boundaries and word-object relations. Adult learners were able to identify both kinds of statistical relations from the same input. Moreover, their learning was actually facilitated by the presence of two simultaneously present relations. Eight-month-old infants, however, do not appear to benefit from the presence of regular relations between words and objects. Adults, like 8-month-olds, did not benefit from regular audio-visual correspondences when they were tested with tones, rather than linguistic input. These differences in learning outcomes across age and input suggest that both developmental and stimulus-based constraints affect statistical learning.     

Modeling the Influence of Language Input Statistics on Children's Speech Production

We trained a computational model (the Chunk-Based Learner; CBL) on a longitudinal corpus of child–caregiver interactions in English to test whether one proposed statistical learning mechanism—backward transitional probability—is able to predict children's speech productions with stable accuracy throughout the first few years of development. We predicted that the model less accurately reconstructs children's speech productions as they grow older because children gradually begin to generate speech using abstracted forms rather than specific “chunks” from their speech environment. To test this idea, we trained the model on both recently encountered and cumulative speech input from a longitudinal child language corpus. We then assessed whether the model could accurately reconstruct children's speech. Controlling for utterance length and the presence of duplicate chunks, we found no evidence that the CBL becomes less accurate in its ability to reconstruct children's speech with age.     

The curse of knowledge: first language knowledge impairs adult learners' use of novel statistics for word segmentation

We investigated whether adult learners' knowledge of phonotactic restrictions on word forms from their first language impacts their ability to use statistical information to segment words in a novel language. Adults were exposed to a speech stream where English phonotactics and phoneme co-occurrence information conflicted. A control where these did not conflict was also run. Participants chose between words defined by novel statistics and words that are phonotactically possible in English, but had much lower phoneme contingencies. Control participants selected words defined by statistics while experimental participants did not. This result held up with increases in exposure and when segmentation was aided by telling participants a word prior to exposure. It was not the case that participants simply preferred English-sounding words, however, when the stimuli contained very short pauses, participants were able to learn the novel words despite the fact that they violated English phonotactics. Results suggest that prior linguistic knowledge can interfere with learners' abilities to segment words from running speech using purely statistical cues at initial exposure.     

Adjacent and Non-Adjacent Word Contexts Both Predict Age of Acquisition of English Words: A Distributional Corpus Analysis of Child-Directed Speech

Children show a remarkable degree of consistency in learning some words earlier than others. What patterns of word usage predict variations among words in age of acquisition? We use distributional analysis of a naturalistic corpus of child-directed speech to create quantitative features representing natural variability in word contexts. We evaluate two sets of features: One set is generated from the distribution of words into frames defined by the two adjacent words. These features primarily encode syntactic aspects of word usage. The other set is generated from non-adjacent co-occurrences between words. These features encode complementary thematic aspects of word usage. Regression models using these distributional features to predict age of acquisition of 656 early-acquired English words indicate that both types of features improve predictions over simpler models based on frequency and appearance in salient or simple utterance contexts. Syntactic features were stronger predictors of children's production than comprehension, whereas thematic features were stronger predictors of comprehension. Overall, earlier acquisition was predicted by features representing frames that select for nouns and verbs, and by thematic content related to food and face-to-face play topics; later acquisition was predicted by features representing frames that select for pronouns and question words, and by content related to narratives and object play.     

Rapid acquisition of phonological alternations by infants

We explore whether infants can learn novel phonological alternations on the basis of distributional information. In Experiment 1, two groups of 12-month-old infants were familiarized with artificial languages whose distributional properties exhibited either stop or fricative voicing alternations. At test, infants in the two exposure groups had different preferences for novel sequences involving voiced and voiceless stops and fricatives, suggesting that each group had internalized a different familiarization alternation. In Experiment 2, 8.5-month-olds exhibited the same patterns of preference. In Experiments 3 and 4, we investigated whether infants' preferences were driven solely by preferences for sequences of high transitional probability. Although 8.5-month-olds in Experiment 3 were sensitive to the relative probabilities of sequences in the familiarization stimuli, only 12-month-olds in Experiment 4 showed evidence of having grouped alternating segments into a single functional category. Taken together, these results suggest a developmental trajectory for the acquisition of phonological alternations using distributional cues in the input.