ERROR-DETECTION, SELF-MONITORING AND SELF-REPAIR IN SPEECH PRODUCTION Stefanie Pillai Dept. of English Language, Faculty of Languages and Linguistics, University of Malaya ABSTRACT: Normal spontaneous speech is characterized by hesitation (silent pauses, filled pauses and prolongations) and self-repairs (repetitions, deletions, substitutions and insertions). The timings of self-repairs have been investigated to explain the processes involved in selfmonitoring and self-repair in the speech production process. Analyses of timings from the onset of an error to the interruption point (error-to-cut-off), and from the interruption point to the onset of a self-repair (cut-off-to-repair) found in the speech of 67 callers to a radio programme, support pre-articulatory monitoring of speech (Levelt, 1983, 1989). There also seems to be evidence of the planning of self-repairs taking place before the interruption point as indicated by a large number of cut-off-to-repair intervals of 0ms. INTRODUCTION Normal spontaneous speech contains hesitation, such as silent pauses, filled pauses and prolongations. There is also often evidence of some form of self-repair in such speech, where words or smaller units are repeated, deleted, substituted or inserted. Such phenomena occur as speakers plan and produce utterances. The fact that speakers hesitate and repair their speech (although the two need not occur together), provide valuable psycholinguistic information about how speakers monitor and repair their own speech within the process of speech production. In relation to this, errorto-cut-off intervals (measured from the onset of an error to the point at which speech is interrupted); and cut-off-to repair intervals (measured from the interruption point to the onset of the repair) have been investigated to explain how the processes of self-monitoring and self-repair work. These time intervals are shown in Figure 1. sometime back 1[in] 2 [ah] 3[near] the ah 1[horse] 2 3[horserace] area Figure 1. The interval between 1 (the error) and 2 (perceived interruption point) is the error-to-cut-off time. The interval between 2 and 3 (the onset of repair) is the cut-off-to repair time. Note that there can be an editing phase between 2 and 3 containing a silent or filled pause. Levelt (1983) found that errors were often interrupted very quickly, even at mid-segment. The implication of such quick interruptions was that the speaker could not have detected the error while attending to his overt speech. Thus, Levelt (1983, 1989) proposed that speakers monitor their inner speech. According to what is known as the main interruption rule, when an error is detected, whether internally or auditorily, speech is immediately interrupted (Nooteboom, 1980; Levelt, 1983). This means that short error-to-cut-off intervals are to be expected. Thus in an incremental model of speech production such as Levelt s, error-detection is followed by the decision to interrupt speech. This in turn is followed by the planning of the repair (repairplanning), which is thought to take place only upon interruption. If this is true, then short cut-off-torepair intervals should not be anticipated. This is contrary to the short cut-off-to-repair intervals found by Blackmer and Mitton (1991), suggesting that repair-planning must have occurred before speech was interrupted. The question then remains as to when repair-planning is initiated. Accepted after abstract review page 533
This paper seeks to examine error-to-cut-off and cut-off-to-repair intervals in naturally-occurring spontaneous speech to gain a better understanding of how error-detection and repair-planning work in self-repairs by answering two questions. First, do the error-to-cut-of intervals suggest prearticulatory monitoring? Second, do the cut-off-to-repair intervals allow for re-planning upon interruption? METHOD The data comprised the speech of 67 callers to an English radio programme. The callers were discussing topics set by the presenters. Only the conversational turns of the callers were analysed. The programmes were audio taped, and then transcribed orthographically. Instances of self-repairs in the transcripts were then coded. For this purpose, a self-repair was defined as the correction of errors without external prompting (Postma, 2000:98). The types of self-repairs that were coded were repetitions, deletions, substitutions and insertions (Lickley, 1998). Examples of such repairs from the data are shown in Figure 2. Types of Self- Repairs Previous Error Editing Phase Repair Continuation Repetitions it s quite em quite correct Deletions we don t have er I mean that is a.. Substitutions the w^ husbands is the general of the households Insertions no but I m what I m saying is Figure 2. Types of self-repairs, showing the error (e.g. phonetic, lexical, syntactic errors or inappropriate use), the editing phase and the repair, as well as what was said before (prior) and after (continuation) the self-repair Any form of hesitation that occurred in the self-repairs was also marked. These included silent pauses, filled pauses (including editing expressions like I mean) and prolonged syllables, which occurred before the interruption point. The parts of the data that contained self-repairs were then digitilised using the Computerized Speech Lab system. Spectrograms of these self-repairs were examined together with their related auditory playback. Following this, error-to-cut-off and cut-off-to-repair intervals were marked on the spectrograms, and measured. The cut-off may be preceded by a prolonged word and/or followed by an editing phase containing filled pauses and/ or silent pauses. RESULTS A total of 239 self-repairs were found in the data. These repairs comprised repetitions (55%), deletions (22%), substitutions (15%) and insertions (8%). Error-to-cut-off intervals The distribution of error-to-cut-off intervals for all the self-repairs had a mean of 347ms, s.d 231ms. The percentage of repairs that had error-to-cut-off intervals below 100ms was 6.7%, while 13.8% were below 150ms. All the four types of repairs had error-to-cut-off intervals below 150ms and 100ms as shown on Table 1. Accepted after abstract review page 534
Table 1. Error-to-cut-off intervals in self-repairs according to repair type. Repairs N Mean s.d. <100ms <150ms (ms) (ms) Repetitions 131 346 219 10.9 14.7 Deletions 53 316 201 7.6 22.2 Substitutions 35 365 195 2.9 8.6 Insertions 20 403 390 10 15 All Self-Repairs 239 348 245 6.7 13.8 Table 2. Comparison of error-to-cut-off intervals found in the present study and in Blackmer and Mitton (1991). In order to compare the data with Blackmer and Mittons (1991) findings, a separate analysis was carried out by leaving out repetitions, which were considered as covert repairs in their study. Although Blackmer and Mitton categorised repairs differently, the percentage of error-to-cut-off intervals below 100ms and 150ms is around the same range as can be seen in Table 2. In other words, although the figures are small in both studies, short error-to-cut-off intervals do occur in selfrepairs. Error-tocut-off Present Study Blackmer and Mitton Mean 348 528 (ms) s.d (ms) 245 300 <100 6.5 5.3 ms <150ms 16.7 14.5 Note. For the purpose of comparison the figures for the present study only consist of deletions, insertions and substitutions, since Blackmer and Mitton s overt repairs did not include repetitions. Cut-off-to-repair intervals The mean duration of cut-off-to repair intervals for self-repairs was 134ms, s.d 261ms. Most of the self-repairs had cut-off-to-repair intervals of less than 250ms (79.4%), with 71.1% of them being less than 100ms. In fact 67% of the self-repairs had cut-off to repair times of 0ms. In other words, the repair immediately followed the offset of the error (see examples in Figure 2). Table 3 shows that at least half or more of all the four types of self-repairs had cut-off-to-repair intervals of 0ms. Table 3. Cut-off-to-repair intervals in self-repairs according to repair type. Repairs N Mean s.d. 0ms <100ms <250ms (ms) (ms) Repetitions 131 86 198 64 74.4 82.4 Deletions 53 200 331 54.7 56.6 69.8 Substitutions 35 148 303 65.7 68.6 74.3 Insertions 20 171 227 50 55 70 All Self-Repairs 239 134 261 67 71.1 79.4 The results were compared to the overt repairs in Blackmer and Mitton (1991). In order to do this, repetitions were once again omitted from the figures. Blackmer and Mitton reported a mean of 332ms, s.d.282 for all their overt repairs, with 48.6% of the overt repairs having cut-off-to-repair Accepted after abstract review page 535
intervals of less than 100ms, and 19.2 % of them with a 0ms interval. A higher percentage was found for the self-repairs (minus repetitions) in this study as can be seen in Table 4. This could be due to the different methods used to categorize repairs in the two studies. What is significant, however is that short cut-off-to-repair intervals do exist in self-repairs. DISCUSSION Table 4. Comparison of cut-off-to-repair intervals found in the present study and in Blackmer and Mitton (1991). Error-tocut-off Present Study Blackmer and Mitton Mean 192 332 (ms) s.d (ms) 313 282 0 ms 57.4 19.2 <100ms 60.2 48.6 Note. For the purpose of comparison the figures for the present study only consist of deletions, insertions and substitutions, since Blackmer and Mitton s overt repairs did not include repetitions. The fact that it can take less than 150ms from the production of an error to the point at which speech is interrupted confirms that there is pre-articulatory monitoring. Given that the time from the detection of an error to the instruction to stop production may take about 200ms (Ladefoged et al, 1973; Levelt, 1983; Logan and Cowan, 1984), it would not be possible for detection to occur after a speaker has produced the error. Thus, error detection is not solely dependent on auditory comprehension. This is also confirmed by the fact that speakers often stop themselves mid-segment (Levelt, 1989; Bear et al., 1992; Nakatani and Hirschberg, 1993; Lickley, 1994; Shriberg, 2001). For example, 67% of selfrepairs that had error-to-cut-off intervals of less than 150ms were stopped mid-segment (see example in Figure 2). Short cut-off-to-repair intervals also have implications planning time of repair. A substantial number of self-repairs had cut-off-to-repair intervals of 0ms, discounting the possibility of repair-planning taking place only upon interruption. This is because the repairs must have been ready for articulation at the moment of interruption. This, in turn, suggests that the repairs must have been planned prior to the cut-off, perhaps even during the production of the error in the case of overt repairs. While Blackmer and Mitton (1991) propose that this may be explained by the fact that planning goes on incrementally while one is speaking, and is stored in a buffer until it is ready for articulation, Harsuiker and Kolk (2001) suggest that both the processes of interruption and repair are triggered at the same time by error detection. This would explain how a repair can be ready at the point of interruption. In other words, interrupting and repair-planning are seen as different but parallel processes. The process of interrupting one s speech is therefore not dependent on the need to make repairs, but can even occur for other reasons (Hartsuiker and Kolk, 2001). SUMMARY AND CONCLUSION This paper shows that consistent with other studies (Blackmer and Mitton, 1991; Harsuiker and Kolk, 2001), error-to-cut-off and cut-off-to-repair intervals can be very short. Whilst short error-to-cut-off intervals confirm that there is pre-articulatory monitoring in speech production as proposed in Levelt s model, there does not seem to be sufficient explanation as to how repair-planning really works given the short cut-off-to-error intervals. Further investigation, such as the computational tests conducted by Hartsuiker and Kolk (2001), is needed to provide a more detailed explanation of the way in which the repair-planning mechanism functions within a model of self-monitoring. Accepted after abstract review page 536
ACKNOWLEDGEMENTS This research is part of the author s PhD research. The author is indebted to her supervisors, Dr Kamila Ghazali (University of Malaya) and Dr Gerry Docherty (University of Newcastle). Part of this research was conducted at the Dept. of Speech, University of Newcastle, supported by a one-year Commonwealth Split-Site Ph.D. Scholarship. REFERENCES Bear, J., Dowding, J and Shriberg E.E. (1992). Integrating multiple knowledge sources for detection and correction of repairs in human-computer dialog, Proceedings of the 30 th Annual Meeting of the Association for Computational Linguistics, 56-63. Blackmer, E.R. and Mitton J.L. (1991). Theories of monitoring and the timing of repairs in spontaneous speech. Cognition 39, 173-194. Harsuiker, R.J. and Kolk, H.J. (2001) Error monitoring in speech production: A computational test of the perceptual loop theory. Cognitive Psychology 42, 113-157. Lackner, J.R. and Tuller, B.H. (1979). Role of efference monitoring in the detection of self-produced speech errors, in W.E. Cooper and E.C.T. Walker (eds.), Sentence Processing, Hillsdale, New Jersey: Erlbaum, pp. 281-294. Ladefoged, P., Silverstein, R. and Papcun, G. (1973) Interruptibility of speech. Journal of the Acoustical Society of America 54, 1105-1108. Levelt, W.J.M. (1983). Monitoring and self-repair in speech, Cognition 14, 41-104. Levelt, W.J.M. (1989). Speaking: From intention to articulation. Cambridge, Massachusetts: MIT Press. Lickley, R.J. (1994). Detecting disfluency in spontaneous speech. Ph.D. dissertation, University of Edinburgh. Lickley, R.J. (1998). HCRC Disfluency Coding Manual, HCRC/TR-100, HCRC, University of Edinburgh at http://www.ling.ed.ac.uk/~robin/maptask/hcrcdsm-02 Logan, G.D. and Cowan, W.B. (1984). On the ability to inhibit thought and action: A theory of an act of control. Psychological Review 91, 295-327. Postma, A. (2000) Detection of errors during speech production; a review of speech monitoring models. Cognition 77 (2), 97-132. Nakatani, C.H. and Hirschberg, J. (1994). A corpus-based study of repair cues in spontaneous speech. Journal of the Acoustical Society of America 95 (3), 1603-1616. Nooteboom, S.G. (1980). Speaking and unspeaking: Detection and correction of phonological and lexical errors in spontaneous speech, in V.A. Fromkin (ed.) Errors in Linguistic Performance: Slips of the Tongue, Ear, Pen and Hand, New York: Academic Press, pp. 87-96. Shriberg, E. (2001) To err is human: Ecology and acoustics of speech disfluencies. Journal of the International Phonetic Association 31 (1), 153-169. Accepted after abstract review page 537