Research ArticleAssessing the distinctiveness of phonological features in word recognition: Prelexical and lexical influences
Introduction
What makes two words sound similar to each other? Consider the English word pin – /pɪn/. Intuitively, we can understand how a word like shin – /ʃɪn/ sounds more similar to pin than a word like train – does. Indeed pin and shin form a minimal pair; the two words are minimally different, in that they share all but one phoneme. Yet cross-modal priming experiments have shown that a word like bin – /bɪn/, which also forms a minimal pair with pin, more strongly activates pin than shin does (e.g., Connine et al., 1993, Milberg et al., 1988). This is because the segments that distinguish pin from shin share fewer phonological features than those that distinguish pin from bin. Now consider the word tin – /tɪn/. Both the /t/ in tin and the /b/ in bin are one feature different from the /p/ in pin (a difference in place and voicing1 respectively). Is the nature of the featural difference pertinent for the notion of similarity?
Research on lexical perception has demonstrated that featural differences in one’s native language are not all perceived as equally distinct. In both English (Cole, Jakimik, & Cooper, 1978) and Dutch (Ernestus & Mak, 2004), mispronunciations have been shown to be less disruptive for word recognition (i.e., easier to recognize) if they involve a change in voicing than if they involve a change in place or in manner. This indicates that a difference in voicing is perceived as less stark than a difference in another major class feature in these languages. More recently, Martin and Peperkamp (2015) exposed French listeners to a series of auditorily- (or audiovisually-) presented nouns supposedly produced by a stroke patient. These included correctly pronounced words, mispronounced words, and non-words that did not resemble any real word. Participants were asked to press a button when they recognized a word – whether it was correctly pronounced or mispronounced – and report it. All mispronunciations involved a change in one of the major class features: voicing, manner, or place on a word-initial obstruent. The results from the audio-only version of that experiment, reported as the proportion of correctly identified mispronounced words, are reproduced in Fig. 1.2 Similar to the previous findings for English and Dutch (Cole et al., 1978, Ernestus and Mak, 2004), words with a voicing mispronunciation were more likely to be recognized than those with a manner or a place mispronunciation. For example, the word sommet, /sɔmɛ/ – “summit” was more likely to be recognized when it was presented as /zɔmɛ/, with a mispronunciation in voicing, than when it was presented as /fɔmɛ/ or /tɔmɛ/ (a place or manner mispronunciation, respectively). Thus, the voicing feature’s role in contrasting words from one another is perceived as different than that of the other features.
The sources of this asymmetry remain unclear, however, and could be due to a number of factors. Most obvious is the acoustic proximity of the sounds being considered. Some sounds are acoustically closer, and will thus be perceived as more similar than other, distant, sounds. A further source of bias is language-specific knowledge. Listeners may use knowledge of their native language for the purposes of efficient word recognition. That is, they may preferentially attend to cues associated with featural contrasts which are more informative in their language. Indeed, listeners are influenced both by acoustic information and by language-specific knowledge (Ernestus and Mak, 2004, Johnson and Babel, 2010). Ernestus and Mak (2004), for example, argued that Dutch listeners are influenced by a process of initial fricative devoicing in their language, which renders voicing information on these segments uninformative. This would explain why these listeners ignore voicing mispronunciations more often than manner mispronunciations in a lexical decision task. Similarly, Johnson and Babel (2010) found language-specific influence using a similarity judgment task. They had native English- and Dutch-speaking participants rate the similarity of pairs of VCV non-words containing English fricatives, and showed that Dutch listeners rated [s], [ʃ], and [θ] as more similar to each other than English listeners did. They argued that this is due to the phonological status of these sounds in the respective languages. While all three sounds are distinctive in English, [ʃ] and [θ] are not phonologically distinctive in Dutch; the former is a contextual allophone of /s/ and the latter does not occur at all. However, in an AX discrimination experiment, Dutch listeners’ response times were not shown to differ from English listeners’; both groups discriminated the same pairs of sounds equally well. The authors argued that their discrimination task reveals low-level acoustic differences between the stimuli, with some of the contrasts yielding longer response times because of their acoustic proximity (e.g., [f]∼[θ] and [h]∼[x]), regardless of the native language of the listener, while their similarity judgment task reveals language-specific influences, with Dutch listeners being perturbed by the absence of [θ] and [ʃ] as phonemes in their native language.
Note, though, that this reasoning does not explain why in English and French, voicing mispronunciations are also harder to detect (Cole et al., 1978, Martin and Peperkamp, 2015), because the voicing feature is fully distinctive in these languages (voicing contrasts can be neutralized in English and French but never word-initially). These results do not necessarily imply that listeners are not influenced by lexical patterns during word recognition. Indeed, following, inter alia, Hall (2013), we argue that a more gradient understanding of “distinctiveness” is necessary to properly address this issue. If, for example, there were fewer voicing minimal pairs than place and manner minimal pairs in English and French, this could explain why words presented with voicing mispronunciations were perceived as closer to the target word. Here, we further explore gradient distinctiveness using a combination of experimental and computational techniques.
The specific aim of our research is to disentangle low-level, prelexical influences from top-down, lexical ones in word recognition. To this end, we take French obstruents as a case study, allowing for a direct comparison with the results on lexical perception from the mispronunciation detection task reported in Martin and Peperkamp (2015), which we take as our starting point. Building on those results, we start off with an examination of the way phonetic differences between features are perceived outside of lexical context, using a prelexical discrimination task. We then examine the French lexicon by measuring the functional load of various feature contrasts as a proxy for the lexical knowledge shared by speakers of French. This allows us to understand if there are asymmetries in the usage of these different features, even though they are not affected by any phonological process. Finally, we compare our results with the word recognition results reported in Martin and Peperkamp (2015), and propose that the relative weight of phonological features during word recognition is determined jointly by the role of these features in both bottom-up acoustic perception and top-down lexical knowledge.
Section snippets
Prelexical perception
The perceptual similarity of speech sounds has been investigated for decades, focusing mostly on the effects of different types of noise on perceptual confusion (e.g., Bell et al., 1989, Cutler et al., 2004, Miller and Nicely, 1955, Weber and Smits, 2003). For instance, Miller and Nicely (1955) presented a series of English syllables embedded in different kinds of noise (including low-pass filtering and white noise) at various signal-to-noise ratios (SNR) and asked participants to report what
Functional load
The term functional load, in a broad sense, refers to the amount of work a phonemic contrast does in a language to distinguish words from one another. Consider the English distinction between /θ/ and /ð/ (the “th” sounds in think and that respectively). Although these sounds are distinctive in English, they actually only disambiguate a handful of words (e.g., ether∼either in American English), and the contrast is therefore considered to have a low functional load. Compare this to the high
General discussion
French listeners are more likely to recognize a mispronounced version of an obstruent-initial word if the mispronunciation concerns the voicing feature than if it concerns the place or manner features (Martin & Peperkamp, 2015). Thus, in French obstruents, both place and manner are more important for word recognition than voicing (at least in nouns), akin with findings in other languages (Cole et al., 1978, Ernestus and Mak, 2004). Where does this asymmetry come from? We examined two sources:
Acknowledgements
This work was supported by ANR-13-APPR-0012 LangLearn, ANR-10-LABX-0087 IEC and ANR-10-IDEX-0001-02 PSL∗. We would like to thank Rory Turnbull for comments and discussion, and Thomas Schatz for his help implementing his model of the ABX task.
References (47)
- et al.
Phoneme similarity and confusability
Journal of Memory and Language
(2005) - et al.
Do the beginnings of spoken words have a special status in auditory word recognition?
Journal of Memory and Language
(1993) - et al.
Hearing speech sounds: Top-down influences on the interface between audition and speech perception
Hearing Research
(2007) Consonant clusters and phonological syllables in French
Lingua
(1995)- et al.
Distinctive phonological features differ in relevance for both spoken and written word recognition
Brain and Language
(2004) - et al.
Perception of wordlikeness: Effects of segment probability and length on the processing of nonwords
Journal of Memory and Language
(2000) - et al.
What makes words sound similar?
Cognition
(2005) - et al.
On the perceptual basis of distinctive features: evidence from the perception of fricatives by Dutch and English speakers
Journal of Phonetics
(2010) - et al.
Bridging phonological system and lexicon: Insights from a corpus study of functional load
Journal of Phonetics
(2015) - et al.
High functional load inhibits phonological contrast loss: a corpus study
Cognition
(2013)
Interactive factors in consonant confusion patterns
The Journal of the Acoustical Society of America
Perceptibility of phonetic features in fluent speech
The Journal of the Acoustical Society of America
Patterns of English phoneme confusions by native and non-native listeners
The Journal of the Acoustical Society of America
Phonological typicality influences on-line sentence comprehension
Proceedings of the National Academy of Sciences of the United States of America
Phonological typicality influences sentence processing in predictive contexts: Reply to Staub, Grant, Clifton, and Rayner (2009)
Journal of Experimental Psychology. Learning, Memory, and Cognition
A probabilistic model of phonological relationships from contrast to allophony
A typology of intermediate phonological relationships
The Linguistic Review
Phonological CorpusTools
Speech perception, well-formedness, and the statistics of the lexicon
Grammatical constraints on phonological encoding in speech production
Psychonomic Bulletin & Review
Cited by (9)
Marginal contrast in loanword phonology: Production and perception
2022, Laboratory Phonology