File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/93/p93-1007_metho.xml
Size: 30,233 bytes
Last Modified: 2025-10-06 14:13:29
<?xml version="1.0" standalone="yes"?> <Paper uid="P93-1007"> <Title>A SPEECH-FIRST MODEL FOR REPAIR DETECTION AND CORRECTION</Title> <Section position="3" start_page="46" end_page="52" type="metho"> <SectionTitle> INTERRUPTION SITE (IS). The DISFLUENCY INTERVAL </SectionTitle> <Paragraph position="0"> (nI) extends from the IS to the resumption of fluent speech, and may contain any combination of silence, pause fillers ('uh', 'urn'), or CUE PHRASES (e.g., 'Oops' 2This is consistent with Levelt (1983)'s observation that the material to be replaced and the correcting material in a repair often share structural properties akin to those shared by coordinated constituents.</Paragraph> <Paragraph position="1"> or 'I mean'), which indicate the speaker's recognition of his/her performance error. The REPAIR INTERVAL corresponds to the utterance of the correcting material, which is intended to 'replace' the reparandum. It extends from the offset of the DI tO the resumption of non-repair speech. In Example (6), for example, the reparandum occurs from 1 to 2, the DI from 2 to 3, and the repair interval from 3 to 4; the Is occurs at 2.</Paragraph> <Paragraph position="2"> (6) Give me airlines 1 \[ flying to Sa- \] 2 \[ SILENCE uh SILENCE \] 3 \[ flying to Boston \] 4 from San Francisco next summer that have business class.</Paragraph> <Paragraph position="3"> RIM provides a framework for testing the extent to which cues from the speech signal contribute to the identification and correction of repair utterances. RIM incorporates two main assumptions of Hindle (1983): (1) correction strategies are linguisticallyrulegoverned, and (2) linguistic cues must be available to signal when a disfluency has occurred and to 'trigger' correction strategies. As Hindle noted, if the processing of disfluencies were not rule-governed, it would be difficult to reconcile the infrequent intrusion of disfluencies on human speech comprehension, especially for language learners, with their frequent rate of occurrence in spontaneous speech. We view Hindle's results as evidence supporting (1). Our study tests (2) by exploring the acoustic and prosodic features of repairs that might serve as a form of edit signal for rule-governed correction strategies.</Paragraph> <Paragraph position="4"> While Labov and Hindle proposed that an acoustic-phonetic cue might exist at precisely the Is, based on our analyses and on recent psychotinguistic experiments (Lickley et al., 1991), this proposal appears too limited. Crucially, in RIM, we extend the notion of edit signal to include any phenomenon which may contribute to the perception of an &quot;abrupt cut-off&quot; of the speech signal -- including cues such as coarticulation phenomena, word fragments, interruption glottalization, pause, and prosodic cues which occur in the vicinity of the disfluency interval. RIM thus acknowledges the edit signal hypothesis, that some aspect of the speech signal may demarcate the computationally key juncture between the reparandum and repair intervals, while extending its possible acoustic and prosodic manifestations.</Paragraph> <Section position="1" start_page="47" end_page="48" type="sub_section"> <SectionTitle> Acoustic-Prosodic Characteristics of Repairs </SectionTitle> <Paragraph position="0"> We studied the acoustic and prosodic correlates of repair events as defined in the RIM framework with the aim of identifying potential cues for automatic repair processing, extending a pilot study reported in (Nakatani and Hirschberg, 1993). Our corpus for the current study consisted of 6,414 utterances produced by 123 speakers from the ARPA Airline Travel and Information System (ATIS) database (MADCOW, 1992) collected at AT&T, BBN, CMU, SRI, and TL 334 (5.2%) of these utterances contain at least one repair~ where repair is defined as the self-correction of one or more phonemes (up to and including sequences of words) in an utterance) Orthographic transcriptions of the utterances were prepared by ARPA contractors according to standardized conventions. The utterances were labeled at Bell Laboratories for word boundaries and intonational prominences and phrasing following Pierrehumbert's description of English intonation (Pierrehumbert, 1980). Also, each of the three RIM intervals and prosodic and acoustic events within those intervals were labeled.</Paragraph> <Paragraph position="1"> Identifying the Reparandum Interval Our acoustic and prosodic analysis of the reparandum interval focuses on acoustic-phonetic properties of word fragments, as well as additional phonetic cues marking the reparandum offset. From the point of view of repair detection and correction, acoustic-prosodic cues to the onset of the reparandum would clearly be useful in the choice of appropriate correction strategy. However, recent perceptual experiments indicate that humans do not detect an oncoming disfluency as early as the onset of the reparandum (Lickley et al., 1991; Lickley and Bard, 1992). Subjects were generally able to detect disfluencies before lexical access of the first word in the repair. However, since only a small number of the test stimuli employed in these experiments contained reparanda ending in word fragments (Lickley et al., 1991), it is not clear how to generalize results to such repairs. In our corpus, 74% of all reparanda end in word fragments. 4 Since the majority of our repairs involve word fragmentation, we analyzed several lexical and acoustic-phonetic properties of fragments for potential use in fragment identification. Table 1 shows the broad word class of the speaker's intended word for each fragment, where the intended word was recoverable. There is a clear tendency for fragmentation at the reparandum offset to occur in content words rather than function words.</Paragraph> <Paragraph position="2"> 3In our pilot study of the SRI and TI utterances only, we found that repairs occurred in 9.1% of utterances (Nakatani and Hirschberg, 1993). This rate is probably more accurate than the 5.2% we find in our current corpus, since repairs for the pilot study were identified from more detailed transcriptions than were available for the larger corpus. Table 2 shows the distribution of fragment repairs by length. 91% of fragments in our corpus are one syllable or less in length. Table 3 shows the distri- null ments (N=288) bution of initial phonemes for all words in the corpus of 6,414 ATIS sentences, and for all fragments, single syllable fragments, and single consonant fragments in repair utterances. From Table 3 we see that single con-</Paragraph> </Section> <Section position="2" start_page="48" end_page="48" type="sub_section"> <SectionTitle> by Fragment Length </SectionTitle> <Paragraph position="0"> sonant fragments occur more than six times as often as fricatives than as stops. However, fricatives and stops occur almost equally as the initial consonant in single syllable fragments. Furthermore, we observe two divergences from the underlying distributions of initial phonemes for all words in the corpus. Vowel-initial words show less tendency and fricative-initial words show a greater tendency to occur as fragments, relative to the underlying distributions for those classes.</Paragraph> <Paragraph position="1"> Two additional acoustic-phonetic cues, glottalization and coarticulation, may help in fragment identification. Bear et al. (1992) note that INTERRUPTION GLO'I~ALIZATION (irregular glottal pulses) sometimes occurs at the reparandum offset. This form of glottalization is acoustically distinct from LARYNGEALIZA-TION (creaky voice), which often occurs at the end of prosodic phrases; GLOTTAL STOPS, which often precede vowel-initial words; and EPENTHETIC GLOTTALtZATtON. In our corpus, 30.2% of reparanda offsets are marked by interruption glottalization. 5 Although interruption glottalization is usually associated with fragments, not all fragments are glottalized. In our database, 62% of fragments are not glottalized, and 9% of glottalized reparanda offsets are not fragments.</Paragraph> <Paragraph position="2"> 5Shriberg et al. (1992) report glottalization on 24 of 25 vowel-final fragments.</Paragraph> <Paragraph position="3"> Also, sonorant endings of fragments in our corpus sometimes exhibit coarticulatory effects of an unrealized subsequent phoneme. When these effects occur with a following pause (see below), they can be used to distinguish fragments from full phrase-final words -- such as 'fli-' from 'fly' in Example (1).</Paragraph> <Paragraph position="4"> To summarize, our corpus shows that most reparanda offsets end in word fragments. These fragments are usually fragments of content words (based upon transcribers' identification of intended words in our corpus), are rarely more than one syllable long, exhibit different distributions of initial phoneme class depending on their length, and are sometimes glottalized and sometimes exhibit coarticulatory effects of missing subsequent phonemes. These findings suggest that it is unlikely that word-based recognition models can be applied directly to the problem of fragment identification. Rather, models for fragment identification might make use of initial phoneme distributions, in combination with information on fragment length and acoustic-phonetic events at the IS. Inquiry into the articulatory bases of several of these properties of self-interrupted speech, such as glottalization and initial phoneme distributions, may further improve the modeling of fragments.</Paragraph> <Paragraph position="5"> Identifying the Disfluency Interval In the RIM model, the D/includes all cue phrases and filled and unfilled pauses from the offset of the reparandum to the onset o.f the repair. The literature contains a number of hypotheses about this interval (cf. (Blackmet and Mitton, 1991). For our corpus, pause fillers or cue words, which have been hypothesized as repair cues, occur within the DI for only 9.8% (332/368) of repairs, and so cannot be relied on for repair detection.</Paragraph> <Paragraph position="6"> Our findings do, however, support a new hypothesis associating fragment repairs and the duration of pause following the IS.</Paragraph> <Paragraph position="7"> Table 4 shows the average duration of 'silent DI'S (those not containing pause fillers or cue words) compared to that of fluent utterance-internal silent pauses for the Tt utterances. Overall, silent DIS are shorter</Paragraph> </Section> <Section position="3" start_page="48" end_page="50" type="sub_section"> <SectionTitle> Fluent Pauses </SectionTitle> <Paragraph position="0"> than fluent pauses (p<.001, tstat=4.60, df=1516). If we analyze repair utterances based on occurrence of fragments, the DI duration for fragment repairs is significantly shorter than for nonfragments (p<.001, tstat=3.36, df=330). The fragment repair DI duration is also significantly shorter than fluent pause intervals (p<.001, tstat=5.05, df=1439), while there is no significant difference between nonfragment DIS and fluent utterances. So, DIS in general appear to be distinct from fluent pauses, and the duration of DIS in fragment repairs might also be exploited to identify these cases as repairs, as well as to distinguish them from nonfragment repairs. Thus, pausal duration may serve as a general acoustic cue for repair detection, particularly for the class of fragment repairs.</Paragraph> <Paragraph position="1"> Identifying the Repair Several influential studies of acoustic-prosodic repair cues have relied upon texical, semantic, and pragmatic definitions of repair types (Levelt and Cutler, 1983; Levelt, 1983). Levelt & Cutler (1983) claim that repairs of erroneous information (ERROR REPAIRS) are marked by increased intonational prominence on the correcting information, while other kinds of repairs, such as additions to descriptions (APPROPRIATENESS REPAIRS), generally are not. We investigated whether the repair interval is marked by special intonational prominence relative to the reparandum for all repairs in our corpus and for these particular classes of repair. To obtain objective measures of relative prominence, we compared absolute f0 and energy in the sonorant center of the last accented lexical item in the reparandum with that of the first accented item in the repair interval. 6 We found a small but reliable increase in f0 from the end of the reparandum to the beginning of the repair (mean--4.1 Hz, p<.01, tstat=2.49, df=327).</Paragraph> <Paragraph position="2"> There was also a small but reliable increase in amplitude across the oI (mean=+l.5 db, p<.001, tstat=6.07, df=327). We analyzed the same phenomena across utterance-internal fluent pauses for the ATIS TI set and found no reliable differences in either f0 or intensity, although this may have been due to the greater variability in the fluent population. And when we compared the f0 and amplitude changes from reparandum to repair with those observed for fluent pauses, we found no significant differences between the two populations.</Paragraph> <Paragraph position="3"> So, while differences in f0 and amplitude exist between the reparandum offset and the repair onset, we conclude that these differences are too small help distinguish repairs from fluent speech. Although it is not entirely straightforward to compare our objective measures of intonational prominence with Levelt and Cutler's perceptual findings, our results provide only weak support for theirs. And while we find small but significant changes in two correlates of intonational prominence, the distributions of change in f0 and energy for our data are unimodal; when we further test subclasses of Levelt and Cutler's error repairs and appropriateness repairs, statistical analysis does not sup6We performed the same analysis for the last and first syllables in the reparandum and repair, respectively, and for normalized f0 and energy; results did not substantially differ from those presented here.</Paragraph> <Paragraph position="4"> port Levelt and Cutler's claim that the former -- and only the former -- group is intonationally 'marked'.</Paragraph> <Paragraph position="5"> Previous studies of disfluency have paid considerable attention to the vicinity of the DI but little to the repair offset. Although we did not find comparative intonationai prominence across the DI tO be a promising cue for repair detection, our RIM analysis uncovered one general intonational cue that may be of use for repair correction, namely the prosodic phrasing of the repair interval. We propose that phrase boundaries at the repair offset can serve to delimit the region over which subsequent correction strategies may operate.</Paragraph> <Paragraph position="6"> We tested the idea that repair interval offsets are intonationally marked by either minor or major prosodic phrase boundaries in two ways. First, we used the phrase prediction procedure reported by Wang & Hirschberg (1992) to estimate whether the phrasing at the repair offset was predictable according to a model of fluent phrasing. 7 Second, we analyzed the syntactic and lexical properties of the first major or minor intonational phrase including all or part of the repair interval to determine whether such phrasal units corresponded to different types of repairs in terms of Hindle's typology. null The first analysis tested the hypothesis that repair interval offsets are intonationally delimited by minor or major prosodic phrase boundaries. We found that the repair offset co-occurs with minor phrase boundaries for 49% of repairs in the TI set. To see whether these boundaries were distinct from those in fluent speech, we compared the phrasing of repair utterances with the phrasing predicted for the corresponding corrected version of the utterance identified by ATIS transcribers.</Paragraph> <Paragraph position="7"> For 40% of all repairs, an observed boundary occurs at the repair offset where one is predicted; and for 33% of all repairs, no boundary is observed where none is predicted. For the remaining 27% of repairs for which predicted phrasing diverged from observed, in 10% of cases a boundary occurred where none was predicted and in 17%, no boundary occurred when one was predicted.</Paragraph> <Paragraph position="8"> In addition to differences at the repair offset, we also found more general differences from predicted phrasing over the entire repair interval, which we hypothesize may be partly understood as follows: Two strong predictors of prosodic phrasing in fluent speech are syntactic constituency (Cooper and Sorenson, 1977; Gee and Grosjean, 1983; Selkirk, 1984), especially the relative inviolability of noun phrases (Wang and Hirschberg, 1992), and the length of prosodic phrases (Gee and Grosjean, 1983; Bachenko 7Wang & Hirschberg use statistical modeling techniques to predict phrasing from a large corpus of labeled ATIS speech; we used a prediction tree that achieves 88.4% accuracy on the ATIS TI corpus using only features whose values could be calculated via automatic text analysis. Results reported here are for prediction on only TI repair utterances.</Paragraph> <Paragraph position="9"> and Fitzpatrick, 1990). On the one hand, we found occurrences of phrase boundaries at repair offsets which occurred within larger NPs, as in Example (7), where it is precisely the noun modifier -- not the entire noun phrase -- which is corrected. 8 (7) Show me all n- \[ round-trip flights \[ from Pittsburgh \[ to Atlanta.</Paragraph> <Paragraph position="10"> We speculate that, by marking off the modifier intonationaily, a speaker may signal that operations relating just this phrase to earlier portions of the utterance can achieve the proper correction of the disfluency. We also found cases of 'lengthened' intonational phrases in repair intervals, as illustrated in the single-phrase reparandum in (8), where the corresponding fluent version of the reparandum is predicted to contain four phrases.</Paragraph> <Paragraph position="11"> (8) What airport is it \[ is located \[ what is the name of the airport located in San Francisco Again, we hypothesize that the role played by this unusually long phrase is the same as that of early phrase boundaries in NPS discussed above. In both cases, the phrase boundary delimits a meaningful unit for subsequent correction strategies. For example, we might understand the multiple repairs in (8) as follows: First the speaker attempts a vP repair, with the repair phrase delimited by a single prosodic phrase 'is located'. Then the initially repaired utterance 'What airport is located' is itself repaired, with the reparadum again delimited by a single prosodic phrase, 'What is the name of the airport located in San Francisco'.</Paragraph> <Paragraph position="12"> In the second analysis of lexical and syntactic properties, we found three major classes of phrasing behaviors, all involving the location of the first phrase boundary after the repair onset: First, for 44% (163/368) of repairs, the repair offset we had initially identified 9 coincides with a phrase boundary, which can thus be said to mark off the repair interval. Of the remaining 205 repairs, more than two-thirds (140/205) have the first phrase boundary after the repair onset at the right edge of a syntactic constituent. We propose that this class of repairs should be identified as constituent repairs, rather than the lexical repairs we had initially hypothesized. For the majority of these constituent repairs (79%, 110/140), the repair interval contains a well-formed syntactic constituent (see Table 5). If the repair interval does not form a syntactic constituent, it is most often an NP-internal repair (77%, 23/30). The third class of repairs includes those in which the first boundary after the repair onset occurs neither at the repair offset nor at the right edge of a syntactic constituent. This class contains surface or lexical be viewed as either constituent or lexical, we preferred the shorter lexical analysis by default.</Paragraph> <Paragraph position="13"> stituent Repairs (N= 110) repairs (where the first phrase boundary in the repair interval delimits a sequence of one or more repeated words), phonetic errors, word insertions, and syntactic reformulations (as in Example (4)). It might be noted here that, in general, repairs involving correction of either verb phrases or verbs are far less common than those involving noun phrases, prepositional phrases, or sentences.</Paragraph> <Paragraph position="14"> We briefly note evidence against one alternative (although not mutually exclusive) hypothesis, that the region to be delimited correction strategies is marked not by a phrase boundary near the repair offset, but by a phrase boundary at the onset of the reparandum. In other words, it may be the reparandum interval, not the repair interval, that is intonationally delimited. However, it is often the case that the last phrase boundary before the IS occurs at the left edge of a major syntactic constituent (42%, (87/205), even though major constituent repairs are about one third as frequent in this corpus (15%, 31/205). In contrast, phrase boundaries occur at the left edge of minor constituents 27% (55/205) of the time, whereas minor constituent repairs make up 39% (79/205) of the subcorpus at hand.</Paragraph> <Paragraph position="15"> We take these figures as general evidence against the outlined alternative hypothesis, establishing that the demarcation repair offset is a more productive goal for repair processing algorithms.</Paragraph> <Paragraph position="16"> Investigation of repair phrasing in other corpora covering a wider variety of genres is needed in order to assess the generality of these findings. For example, 35% (8/23) of NP-internal constituent repairs occurred within cardinal compounds, which are prevalent in the nTIS corpus due to its domain. The preponderance of temporal and locative prepositional phrases may also be attributed to the nature of the task and domain. Nonetheless, the fact that repair offsets in our corpus are marked by intonational phrase boundaries in such a large percentage of cases (82.3%, 303/368), suggests that this is a possibility worth pursuing.</Paragraph> <Paragraph position="17"> Predicting Repairs from Acoustic and</Paragraph> </Section> <Section position="4" start_page="50" end_page="52" type="sub_section"> <SectionTitle> Prosodic Cues </SectionTitle> <Paragraph position="0"> Despite the small size of our sample and the possibly limited generality of our corpus, we were interested to see how well the characterization of repairs derived from RIM analysis of the ATIS COrpUS would transfer to a predictive model for repairs in that domain. We examined 374 ATIS repair utterances, including the 334 upon which the descriptive study presented above was based. We used the 172 TI and SRI repair utterances from our earlier pilot study (Nakatani and Hirschberg, 1993) as training date; these served a similar purpose in the descriptive analysis presented above. We then tested on the additional 202 repair utterances, which contained 223 repair instances. In our predictions we attemped to distinguish repair Is from fluent phrase boundaries (collapsing major and minor boundaries), non-repair disfluencies, 1deg and simple word boundaries. We considered every word boundary to be a potential repair site. 11 Data points are represented below as ordered pairs <wl,wj >, where wi represents the lexical item to the left of the potential IS and wj represents that on the right.</Paragraph> <Paragraph position="1"> For each <wi,wj >, we examined the following features as potential Is predictors: (a) duration of pause between wi and wj; (b) occurrence of a word fragment(s) within <w~,wj >; (c) occurrence of a filled pause in <wi,wj >; (d) amplitude (energy) peak within wi, both absolute and normalized for the utterance; (e) amplitude of wi relative to wi-i and to wj; (f) absolute and normalized f0 of wi; (g) f0 of wi relative to wi-i and to wj; and (h) whether or not wi was accented, deaccented, or deaccented and cliticized. We also simulated some simple pattern matching strategies, to try to determine how acoustic-prosodic cues might interact with lexical cues in repair identification. To this end, we looked at (i) the distance in words of wi from the beginning and end of the utterance; (j) the total number of words in the utterance; and (k) whether wi or wi-1 recurred in the utterance within a window of three words after wi. We were unable to test all the acoustic-prosodic features we examined in our descriptive analysis, since features such as glottalization and coarticulatory effects had not been labeled in our data base for locations other than DIs. Also, we used fairly crude measures to approximate features such as change in f0 and amplitude, since these .too had been precisely labeled in our corpus only for repair locations and not for fluent speech./2 We trained prediction trees, using Classification and Regression Tree (CART) techniques (Brieman et al., 1984), on our 172-utterance training set. We first included all our potential identifiers as possible predictors. The resulting (automatically generated) decision tree was then used to predict IS locations in our 202ldegThese had been marked independently of our study and including all events with some phonetic indicator of disfluency which was not involved in a self-repair, such as hesitations marked with audible breath or sharp cut-off.</Paragraph> <Paragraph position="2"> llWe also included utterance-final boundaries as data points.</Paragraph> <Paragraph position="3"> 12We used uniform measures for prediction, however, for both repair sites and fluent regions.</Paragraph> <Paragraph position="4"> utterance test set. This procedure identified 186 of the 223 repairs correctly, while predicting 12 false positives and omitting 37 true repairs, for a recall of 83.4% and precision of 93.9%. Fully 177 of the correctly identified ISS were identified via presence of word fragments as well as duration of pause in the DL Repairs not containing fragments were identified from lexical matching plus pausal duration in the DI.</Paragraph> <Paragraph position="5"> Since the automatic identification of word fragments from speech is an unsolved problem, we next omitted the fragment feature and tried the prediction again. The best prediction tree, tested on the same 202-utterance test set, succeeded in identifying 174 of repairs correctly-- in the absence of fragment information- with 21 false positives and 49 omissions (78.1% recall, 89.2% precision). The correctly identified repairs were all characterized by constraints on duration of pause in the DI. Some were further identified via presence of lexical match to the right of wi within the window of three described above, and word position within utterance. Those repairs in which no lexical match was identified were characterized by lower amplitude of wi relative to wj and cliticization or deaccenting of wi. Still other repairs were characterized by more complex series of lexical and acoustic-prosodic constraints.</Paragraph> <Paragraph position="6"> These results are, of course, very preliminary.</Paragraph> <Paragraph position="7"> Larger corpora must certainly be examined and more sophisticated versions of the crude measures we have used should be employed. However, as a first approximation to the characterization of repairs via both acoustic-prosodic and lexical cues, we find these resuits encouraging. In particular, our ability to identify repair sites successfully without relying upon the identification of fragments as such seems promising, although our analysis of fragments suggests that there may indeed be ways of identifying fragment repairs, via their relatively short DI, for example. Also, the combination of general acoustic-prosodic constraints with lexical pattern matching techniques as a strategy for repair identification appears to gain some support from our predictions. Further work on prediction modeling may suggest ways of combining these lexical and acoustic-prosodic cues for repair processing.</Paragraph> <Paragraph position="8"> In this paper, we have presented a&quot;speech-first&quot; model, the Repair Interval Model, for studying repairs in spontaneous speech. This model divides the repair event into a reparandum interval, a disfluency interval, and a repair interval. We have presented empirical results from acoustic-phonetic and prosodic analysis of a corpus of repairs in spontaneous speech, indicating that reparanda offsets end in word fragments, usually of (intended) content words, and that these fragments tend to be quite short and to exhibit particular acoustic-phonetic characteristics. We found that the disfluency interval can be distinguished from intonational phrase boundaries in fluent speech in terms of duration of pause, and that fragment and nonfragment repairs can also be distinguished from one another in terms of the duration of the disfluency interval. For our corpus, repair onsets can be distinguished from reparandum offsets by small but reliable differences in f0 and amplitude, and repair intervals differ from fluent speech in their characteristic prosodic phrasing. We tested our results by developing predictive models for repairs in the ATIS domain, using CART analysis; the best performing prediction strategies, trained on a subset of our data, identified repairs in the remaining utterances with recall of 78-83% and precision of 89-93%, depending upon features examined.</Paragraph> </Section> </Section> <Section position="4" start_page="52" end_page="52" type="metho"> <SectionTitle> Acknowledgments </SectionTitle> <Paragraph position="0"> We thank John Bear, Barbara Grosz, Don Hindle, Chin Hui Lee, Robin Lickley, Andrej Ljolje, Jan van Santen, Stuart Shieber, and Liz Shriberg for advice and useful comments. CART analysis employed software written by Daryl Pregibon and Michael Riley. Speech analysis was done with Entropic Research Laboratory's</Paragraph> </Section> class="xml-element"></Paper>