William D. Raymond

The effect of language model probability on pronunciation reduction

We investigate how the probability of a word affects its pronunciation. We examined 5618 tokens of the 10 most frequent (function) words in Switchboard and 2042 tokens of content words whose lexical form ends in a t or d. Our observations were drawn from the phonetically hand-transcribed subset of the Switchboard corpus, enabling us to code each word with its pronunciation and duration. Using linear and logistic regression to control for contextual factors, we show that words which have a high unigram, bigram, or reverse bigram (given the following word) probability are shorter, more likely to have a reduced vowel, and more likely to have a deleted final t or d. These results suggest that pronunciation models in speech recognition and synthesis should take into account word probability given both the previous and following words, for both content and function words.

Strategy Selection and Use During Classification Skill Acquisition

Two experiments examined 3 variables affecting accuracy, response time, and reports of strategy use in a binary classification skill task. In Experiment 1, higher rule cue salience, allowing faster rule application, produced higher aggregate rule use than lower rule cue salience. After participants were pretrained on the relevant classification rule, rule reports were high but generally declined across training trials; after participants were pretrained on an irrelevant rule, reports of the relevant rule increased across training trials. In Experiment 2, no rule pretraining produced a pattern of results like that obtained with irrelevant rule pretraining in Experiment 1. Presenting novel stimuli during training in Experiment 2 elevated aggregate rule reports relative to conditions where they were absent. Two participant subgroups were identified: those persisting in rule reports and those transitioning from rule to memory reports during training. The proportion of persistent rule users was higher after rule discovery than after relevant rule pretraining. Overall, the results indicate that differences among prior experiments can be reconciled. Further, they raise questions about the inevitability of memory-based automaticity in binary classification, favoring instead strategy choice based on the costs and benefits of a particular strategy and of a shift from one strategy to another.

Linguistic knowledge and language performance in English article variant preference

The knowledge that speakers use in language performance has traditionally been described in terms of abstract rules. However, speaker and listener performance cannot always be captured using a single rule. Alternative models include the interaction of multiple rules, knowledge of association pairs, and generalisation over word similarities. The study tests views of linguistic knowledge against the results of three experiments using English definite and indefinite article variant preference in both a production and a perception task. The results argue against a rule-based model of speaker knowledge or knowledge of article--word combinations, but support the view that performance is based on generalisations sensitive to phonetic, prosodic, and orthographic word features. In addition to initial segment class, definite article variant preference was influenced by a tendency for language users to prefer an alternating rhythm of stressed and unstressed syllables. Effects of sound-letter correspondence were obtained even though participants saw no written materials. The consistency of the effects across modalities argues that the same type of language knowledge underlies both the perception and production of language.

Frequency and predictability effects on the duration of content words in conversation

This study carefully examines the proposal that word forms in conversation are shorter when they have higher probability, where this includes all factors of their context (Jurafsky et al., Frequency and the Emergence of Linguistic Structure, edited by Bybee and Hopper, 2001, pp. 229–254). Higher word frequencies have long been known to be linked to shorter durations. But it has been hard to confirm a direct relationship between the two. One confound is with predictability or conditional probability, which also affects surface form and is closely related to frequency. Another difficulty is controlling factors related both to duration and contextual probability. Linear regression (and sampling to avoid violating independence assumptions) was used to control for phonological form, speech rate, prosodic prominence, previous uses, and neighboring disfluencies. Both greater word frequencies and greater conditional probabilities given the following word have a strong shortening effect on durations of content wor...

Cumulative context effects and variant lexical representations: Word use and English final t/d deletion

Word production variability is widespread in speech, and rates of variant production correlate with many factors. Recent research suggests mental representation of both canonical word forms and distinct reduced variants, and that production and processing are sensitive to variant frequency. What factors lead to frequency-weighted variant representations? An experiment manipulated following context and word repetition for final t/d words in read, narrative English speech. Modeling the experimentally generated data statistically showed higher final-segment deletion in tokens followed by consonant-initial words, but no evidence of increased deletion with repetition, regardless of context. Deletion rates were also higher the greater a words cumulative exposure to consonant contexts (measured from distributional statistics), but there was no effect of word frequency. Token effects are interpreted in terms of articulation processes. The type-level context effect is interpreted within exemplar and usage-based models of language to suggest that experiences with word variants in contexts register as frequency-weighted representations.

Acquisition of morphological variation: The case of the English definite article

Morphological systems have been pivotal in exploring cognitive mechanisms of language use and acquisition. Adult English definite article form preference seems to depend non-deterministically on multiple factors. A corpus study of adult spontaneous speech revealed similar patterns of variability. In an experiment, article variant preferences of three age groups were compared. Children were sensitive to the same phonological factors as adults, but showed effects of more limited experience with articulation and orthography. Preferences across age groups suggest developmental changes, but no evidence that children initially use a default form. Corpus studies of childrens and adults’ speech also revealed no evidence for a default. The results point to overgeneralisation of both article variants, resulting from extended competition between variant forms.

Pairing Words with Syntactic Frames: Syntax, Semantics, and Count-Mass Usage

Two experiments examined English speakers’ choices of count or mass compatible frames for nouns varying in imageability (concrete, abstract) and noun class (count, mass). Pairing preferences with equative (much/many) and non-equative (less/fewer) constructions were compared for groups of teenagers, young adults, and older adults. Deviations from normative usage were, for all ages, larger for count than for mass nouns, for the non-equative than for the equative construction, and for abstract count than for concrete count nouns. These results indicate that mass syntax is not a developmental default, support proposals that mass syntax is more flexible than count syntax, verify the non-prescriptive use of less with count nouns, and extend the interaction of syntax and semantics in noun classification to older ages, with older adults showing a reduced reliance on semantics. Knowledge of frame compatibility and knowledge of noun class are also shown to be largely independent.

Frequency effects in malpropisms

Malapropisms are lexical substitution errors resulting from a failure at the stage of accessing the phonological form corresponding to a lemma—a semantically and structurally specified lexical entry [Fay and Cutler (1977); Garrett (1980)]. An example is tentative fortenable . Since they occur naturally in utterance contexts, malapropisms are an important source of information about this stage of lexical access. A study of over 300 malapropisms confirmed prior findings that errors resemble targets closely in phonological form and in syntactic category. In addition, it was found that errors resemble targets in derivational morphology (independently of phonological similarity), and that there is a strong correlation (r2=0.34) of the text frequencies of the error and target words. The partial correlations of frequency with word length, syntactic category, and segmental similarity only account for one‐half of this correlation (residual r2 about 0.17). Models which treat frequency effects as biases on the decis...

An introduction to the Buckeye corpus of conversational speech

Words in connected speech are subject to many types of phonological variation. Variation is both widespread and conditioned by an array of factors, including speech rate and context. However, little is known about the frequencies of variants or their acoustic realization. The variation‐in‐conversation (ViC) group at Ohio State has set about to quantify and measure the distributional properties of variants, with an eye toward their implications for spoken word perception. The current study reports on a corpus of approximately 40 h of conversation‐style speech recorded during interviews of 43 midwestern speakers, stratified for age and sex. The speech was paired with text transcription, and is being phonetically transcribed using a combination of automatic and hand alignment. Details of the collection and transcription methodology will be presented, along with some parameters of variation that have emerged from the data. For example, in the corpus t is more likely to flap (16% of all underlying t’s) than d ...