Signal-driven computations in speech processing

Science

Volumn 298, Issue 5593, 2002, Pages 604-607

  Peña, Marcela ^a   Bonatti, Luca L ^a   Nespor, Marina ^a   Mehler, Jacques ^a  

^a SISSA   (Italy)

Author keywords

[No Author keywords available]

Indexed keywords

CALCULATIONS; FORMAL LANGUAGES;

SPEECH SIGNALS;

SPEECH;

LANGUAGE; LEARNING; SIGNAL;

ARTICLE; CHILD; CHILD DEVELOPMENT; HUMAN; INFANT; LANGUAGE DEVELOPMENT; LEARNING; PRIORITY JOURNAL; SIGNAL PROCESSING; SPEECH; STATISTICS;

ADULT; CUES; FRANCE; HUMANS; LANGUAGE; LEARNING; LINGUISTICS; PHONETICS; PROBABILITY; SPEECH PERCEPTION; STATISTICS; VOCABULARY;

EID: 0037131255     PISSN: 00368075     EISSN: None     Source Type: Journal    
DOI: 10.1126/science.1072901     Document Type: Article

References (33)

1. A. Christophe, E. Dupoux, J. Bertoncini, J. Mehler, J. Acoust. Soc. Am. 95, 1570 (1994).
2. R. N. Aslin, J. Z. Woodward, N. P. LaMendola, T. G. Bever, in Signal to Syntax: Bootstrapping from speech to grammar in early acquisition, J. Morgan, K. Demuth, Eds, (Erlbaum, Hillsdale, NJ, 1996), pp. 117-134.
3. J. R. Hayes, H. H. Clark, in Cognition and the Development of Language, J. R. Hayes, Ed., (Wiley, New York, 1970), pp. 221-234.
4. J. R. Saffran, E. L. Newport, R. N. Aslin, J. Mem. Lang. 35, 606 (1996).
5. J. R. Saffran, R. N. Aslin, E. L. Newport, Science 274, 1926 (1996).
6. Saffran, Aslin, and Newport (5) use continuous streams of artificial consonant-vowel syllables. All syllables have the same duration, loudness, and pitch. Range of transitional probabilities used with adults is 0.25 to 1.00 within trisyllabic items and 0.05 to 0.60 between. Range of transitionat probabilities in infant studies is 1.00 within and 0.33 between trisyllabic items.
7. J. R. Saffran, E. K. Johnson, R. N. Aslin, E. L. Newport, Cognition 70, 27 (1999).
8. J. Fiser, R. N. Aslin, Psychol. Sci. 12, 499 (2001).
9. _, J. Exp. Psychol. 28, 458 (2002).
10. N. Z. Kirkham, J. A Slemmer, S. P. Johnson, Cognition 83, B35 (2002).
11. G. F. Marcus, S. Vijayan, S. B. Rao, P. M. Vishton, Science 283, 77 (1999).
12. Marcus et al. (11) used lists of trisyllabic items (all consonant-vowel syllables) with either an ABA structure (e.g., GATIGA, LINALI, etc.) or an ABB structure (e.g., GATITI, LINANA, etc.). Each syllable was separated by 250-ms gaps, and each trisyllabic item was presented after a 1-s pause. Infants prefer new items consistent with the familiarization structure to new items inconsistent with it.
13. E. L. Newport, R. N. Aslin, in Proceedings of the 24th Annual Boston University Conference on Language Development, S. C. Howell, S. A. Fish, T. Keith-Lucas, Eds., November 1999, Boston, MA (Cascadilla Press, Sommerville, MA, 2000), pp. 1-21.
14. In a first study using a very similar procedure as Saffran et al. (5), we demonstrated that when adjacent transitional probabilities are manipulated, French participants (n = 12) behave like the English participants they tested. We showed that after 21 min of familiarization, 10 out of 12 participants significantly chose words over part words. The within-word transitional probability range was 0.25 to 1.0 and the between-word transitional probability range was 0.05 to 0.6.
15. 3 is: [talidu], [taRadu], [tafodu]. The part words are either formed by the last syllable of one word and the first two syllables of the following word, i.e., structure CAX ([kitaRa], [kitafo], [gapufo], and [dubeRa]), or by the last two syllables of one word and the first syllable of the following word, i.e., structure XCA ([likita], [lidube], [Radube], [Ragapu), [fogapu]). All words and part words are meaningless in French. Materials and methods, as well as a sample of the familiarization sound file, are available as supporting material on Science Online.
16. Because, in the stream, part words occur less frequently than words, the outcome of experiment 1 might be due to absolute frequency of syllable cooccurrence, rather than to distant transitional probabilities. To disentangle the two explanations, we ran a control experiment in which a 10-min stream was constructed in such a way that each family of words contained one high-frequency item and two items occurring half as frequently. We then tested the two low-frequency words of each family against part words having exactly the same frequency of occurrence in the stream. Adult participants (n = 14) were exposed to the 10-min stream and were tested with 24 word-part word couples equalized in frequency. Even in this case, 70.5% of participants' choices were for the words (P < 0.03). Thus it is not the absolute frequency of occurrence of the trisyllabic items that directs participants' choices, but the distant transitional probability relations among their syllables.
17. Although the stream was synthesized to be monotonous, continuous, and composed of words and part words of equal duration, it is possible that listeners do not perceive the signal as such. The perceptual centers of (monosyllabic and bisyllabic) words in a noncontinuous stream do not correspond to their physical centers [J. Morton, S. Marcus, C. Frankish, Psychol. Rev. 83, 405 (1976)]. A sense of rhythm might arise from an uneven distribution of the perceptual centers of the syllables composing words and part words in the stream used in experiment 1. This might have induced participants to favor words over part words in the test phase irrespective of the transitional probabilities between the syllables composing them. To rule out this hypothesis, we ran a control experiment. We constructed a new familiarization stream transforming most of the part words of experiment 1 into the words of the control, and most of the words of experiment 1 into the part words of the control (by manipulating the nonadjacent syllable transitional probabilities in the familiarization stream). Thus, we chose part words and words so that their average perceptual center distributions, as well as other phonological properties, would become the mirror image of those in experiment 1. We reasoned that if the result of experiment 1 were due to a difference in average perceptual center distributions of words and part words, in this control participants should now prefer the part words over the words. The outcome of the control experiment showed that participants (n = 14) still preferred words (as defined by their nonadjacent transitional probabilities) over part words [mean preference for words = 58%, t(13) = 2.7, P < 0.02]. Thus, perceptual grouping based on syllable perceptual centers or other phonological properties specific to the words and part words we selected in experiment 1 accounts neither for the result of experiment 1 nor for the results of the following experiments reported in this paper.
18. i]. For instance, if the stream contained [puliki]. [puRaki], and [pufoki] as one family and [beliga] in another family, we tested whether participants preferred [pubeki], which had not appeared in the stream, to the part word [likibe] that did appear in it. The syllable [be] appears in the stream, but not as a middle syllable in the family [puXki].
19. The difference between performance in experiments 1 and 2 is significant [t(26) = 3.625, P < 0.001].
20. N. Chomsky, Syntactic Structures (The Hague, Mouton, 1957).
21. S. Pinker, Science 253, 530 (1991).
22. To make sure that the 25-ms gaps remained at subthreshold level, we ran a control experiment. Participants (n = 14) were asked to judge two sequences of synthetic speech drawn from the familiarization used in experiments 2 and 3: one of them had silent gaps at the end of words while the other one was continuous. Both sequences lasted 1 min. Participants were told that they would have to listen to two sequences of an imaginary language and were informed that they would have to answer questions about these speech streams at the end of the presentation. The first question was whether they had heard any silent gaps. Thirteen out of 14 participants responded negatively. Next, all participants were informed that only one of the sequences had gaps and were asked to identify which of the two sequences contained them. Participants responded at chance. Therefore the 25-ms gaps are processed at some level, but participants do not become aware of their presence. A sample of the familiarization sound file is available as supporting material on Science Online.
23. The difference between performance in experiments 2 and 3 is significant [t(26) = -3.087, P < 0.005].
24. J. L. Morgan, E. L. Newport, J. Verb. Learn. Verb. Behav. 20, 67 (1981).
25. J. L. Morgan, R. P. Meier, E. L. Newport, Cognit. Psychol. 19, 498 (1987).
26. _, J. Mem. Lang. 28, 360 (1989).
27. Participants might have included the gaps as separate elements for computing transitional probabilities. As a result, they may have preferred rule words, not because they extracted the structure of the stream, but because they computed probabilities over syllables, pauses, and absence of pauses in the stream and the test items. Thus participants may have analyzed the rule words in the test as having the structure #A1X*C1# and the part words as having the structure #XC2@A3# (where # indicates a pause and @ the absence of a pause). In this case, the transitional probabilities between adjacent elements would favor rule words over part words and no sensitivity to the structure of the rule words would be needed to prefer rule words. This hypothesis makes a prediction that has not been confirmed in a control experiment. Though in experiment 3 the test items do not contain pauses, in this control experiment we tested participants (n = 14) with items including the pauses. Thus, participants compared rule words with structure #A1X*C1# to part words with structure #XC2#A3#. In this case, the presence of the pause in the part words makes the transitional probability of the part word higher than that of the rule word. Therefore, if pauses counted as separate events in the computation, participants should favor the part words over the rule words. Nevertheless, contrary to this prediction, participants still preferred rule words to part words [61.5%, t(13)= 2.8, P < 0.015].
28. G. F. Marcus, The Algebraic Mind: Integrating connectionism and cognitive science (MIT Press, Cambridge, MA, 2001).
29. i will follow at position 3." To evaluate these possibilities, we ran a control experiment. In it, participants were familiarized with the 2-min stream of experiment 5, but in the test phase they listened to pairs of quadrisyllabic items obtained from the words and the part words of experiment 5 by substituting the middle syllable with two syllables that never occurred in the stream. Participants (n = 14) still preferred the item conforming to the structural generalization [mean preference for rule words = 63%, t(13) = 5.012, P = 0.0002]. This shows that participants did not respond on the basis of strict temporal or sequential templates derived from the familiarization stream.
30. L. R. Gleitman, E. Wanner, in Language Acquisition: The State of the Art, L. R. Gleitman, E. Wanner, Eds. (Cambridge Univ. Press, Cambridge, 1982), pp. 3-48.
31. We cannot exclude that longer familiarization with continuous streams might also yield generalizations, but the underlying computations would be different from the rapid projection of generalization that we found in experiments 3 and 5.
32. M. S. Seidenberg, Science 275, 1599 (1997).
33. This research has been supported by a grant from the Regione of Friuli-Venezia-Giulia. We are grateful to D. Aslin, A. Caramazza, S. Dehaene, J. Fodor, J. Morton, G. Marcus, L. Newport, B. Scholl, M. Shukla, and T. Shallice for discussions and suggestions.