Effects of the lexicon and context on speech perceptionLexical StatisticsLexical StatisticsOutlineEffects of lexical properties on word recognitionNeighborhood density/Relative frequencyNeighborhood densityLuce, Pisoni & Goldinger (1990)Luce, Pisoni & Goldinger (1990)Luce, Pisoni & Goldinger (1990)ResultsLuce, Pisoni & Goldinger (1990)A Bayesian model of word recognitionBayes’ TheoremApplication of Bayes’ TheoremApplication of Bayes’ TheoremA Bayesian model of the listener - word frequencyA Bayesian model of the listener - neighborhood densityA Bayesian model of the listener - context effectsA Bayesian model of the listener - context effectsBoothroyd & Nittrouer 1988A Bayesian model of the listener - context effectsInteractions between context and lexical statisticsInteractions between context and lexical statisticsInteractions between context and lexical properties: Neighborhood density/ContextSommers & Danielson (1999)Interactions between context and lexical properties: Neighborhood density/ContextReferencesReferencesReferencesReferencesA Bayesian model of the listener - context effectsMIT OpenCourseWare http://ocw.mit.edu24.910 Topics in Linguistic Theory: Laboratory PhonologySpring 2007 For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.Effects of the lexicon and context on speech perceptionLexical StatisticsThe pronunciation of words does not just depend on their phonological representations (features etc), also• Prosody (phrasing, accentuation)• Speech rate• ‘Lexical’ statistics:– word frequency– neighborhood density• Contextual predictability (cloze probability)Lexical Statistics• Word recognition is also affected by these properties• It has been hypothesized that the production and perception effects are linked.– Words that are more difficult to recognize are pronounced more clearly.Outline• Review some effects of lexical statistics on word recognition.• Present an analysis of these effects in terms of a Bayesian model of word recognition.• Explore predictions concerning interactions between frequency/neighborhood density and contextual predictability.– These effects should be less important where contextual information is available.• Next time: look at corresponding production effects, and general evidence for ‘listener-oriented behavior’ on the part of speakers.Effects of lexical properties on word recognition• Frequency: more frequent words are identified more rapidly and accurately (e.g. Goldinger et al 1996)• Luce (1986) demonstrated that word frequency alone is not a very good predictor of difficulty in word recognition - neglects competition effects.• Recognizing a word involves picking out that word from all of the words in the lexicon.• This process of discrimination may be impeded where there are many words that are perceptually similar to the target word - lexical neighbors.• High frequency facilitates the recognition of the target words, but high frequency neighbors impede recognition.Neighborhood density/Relative frequencyFigure by MIT OpenCourseWare. Adapted from Lindblom 1990.Auditory Similarity(1- Dim Projection)( 2-Dim Projection)Auditory Similarity SpaceStim Word 1 Stim Word 2 Stim Word 3NeighborsFrequency of OccurrenceNeighborhood density• High density: cat: coat, at, scat, cap…• Low density: choice: voice, chase• Also matters how frequent those neighbors areFigure by MIT OpenCourseWare.Vowel High frequency / High frequency / Low frequency / Low frequency /high density low density high density low densityagot dock dot mopalock rock knock sockapot top cot copbad bag dad dabsad sang fad saghalf laugh mash rasheget death debt deafebet check pet pepeIsave gave cage batheeIgame gain dame babeeItape shape cake napeæææLuce, Pisoni & Goldinger (1990)• Tested effects of lexical neighborhood on speed and accuracy of identification of CVC words in noise.• Neighborhood probability rule– p(stimulus word) is probability of correctly identifying the segments of the stimulus.– p(neighborj) is probability of misidentifying the stimulus as (having the segments of) neighbor j.p(ID) =p(stimulus word) × freqsp(stimulus word) × freqs+ p(neighborj) × freqj{}j=1n∑Luce, Pisoni & Goldinger (1990)Predictions: • words with higher frequency should be more accurately identified.• Words with higher stimulus probability (made up of less confusable segments) should be more accurately identified.• Words with more similar neighbors should be less accurately identified.• Words with more high frequency neighbors should be less accurately identified.p(ID) =p(stimulus word) × freqsp(stimulus word) × freqs+ p(neighborj) × freqj{}j=1n∑Luce, Pisoni & Goldinger (1990)• Stimuli: 400 CVC words divided into 8 classes, fully crossing:– High vs. low word frequency– High vs. low stimulus probability– High vs. low frequency-weighted neighborhood probability• Words mixed with white noise (SNR +5dB) and presented to subjects for identification.• Stimulus/neighbor probabilities were estimated from confusion matrices for CV and VC syllables in noise.– Assume confusion probability depends only on position.– p(kɪd|kæt) = p(kons|kons)×p(ɪ|æ)×p(dcoda| tcoda)–p(∅|seg) and p(seg| ∅) were used to for CCVC, CV etc.• Only familiar monosyllabic words were considered.Results• High stimulus probability words identified more accurately than low stimulus probability words.• Words with high frequency-weighted neighborhood probabilities identified less accurately.• High frequency words identified more accurately than low frequency words, but high freq words in dense neighborhoods identified less accurately than low freq words in sparse neighborhoods.Figure by MIT OpenCourseWare. Low HighPercent correctFrequency-weighted neighborhood probability0102030405060708090100high stimulus probhigh stimulus problow stimulus problow stimulus probLow HighPercent correctHigh frequency words Low frequency wordsFrequency-weighted neighborhood probability0102030405060708090100Luce, Pisoni & Goldinger (1990)• Lexical decision: Spoken word or nonword presented.– Subject must decide whether the stimulus is a word or not.• Reaction time to nonword stimuli were slower where:– Mean frequency of neighbors is higher.– Density of neighborhood is higher.– No interaction.• Here neighbors of a word are taken to be all words that can be created from that word by