Phonetic Reduction Research Articles

Phonetic reduction versus phonological deletion of French schwa: Some methodological issues

A categorical phonological process of deletion is traditionally assumed to account for the alternation of schwa with zero in French. This process is assumed to result in two discrete outputs: forms with schwa (i.e., schwa variants) and forms without schwa (i.e., non-schwa variants). However, the two studies we present here suggest a more complex picture. In the first study, we investigate the phonetic variability of schwa in a large number of occurrences of schwa variants and find that schwa, like other segments in French, undergoes phonetic reduction. As a consequence, some tokens without schwa in connected speech may be the result of a process of gradual phonetic reduction rather than the result of a categorical process of alternation. In the second study, we examine the perception of schwa word tokens extracted from connected speech. We show that deciding whether a token was produced with or without the schwa is not always possible. Furthermore, listeners rely on other types of cues than acoustic ones in order to make their judgements (i.e., speech rate, word length and segmental context). These findings have important theoretical and methodological implications that must be taken into account in the empirical study of French schwa alternation.

Phonetic reduction and informational redundancy in self-initiated self-repair in Dutch

This paper presents a phonetic analysis of instances of self-initiated self-repair taken from a corpus of Dutch conversation, focusing on their temporal and articulatory characteristics. In particular, the paper investigates whether error repairs and appropriateness repairs, two subtypes of self-repair which differ in the degree of informational redundancy of the repair stretch, are different in terms of the phonetic relationship between reparandum and repair stretch. One might expect error repairs, whose repair stretch is more informative than that of appropriateness repairs, to be associated with local hyperarticulation in the repair stretch, while appropriateness repairs are more likely to be associated with phonetic reduction. Furthermore, the frequent presence of informationally redundant editing terms and repeated lexical items is expected to result in variation within the repair stretch along a hypo–hyper continuum. The paper reports on temporal measurements with controls for the internal structure of the repair, as well as auditory analysis and an attempt at quantifying a ‘reduction quotient’ for the stretches involved. The results show that local hyperarticulation is rare, and increased phonetic reduction in the repair stretch widespread. No significant differences between error and appropriateness repairs are observed. The findings are discussed with reference to the role of information status and pragmatic constraints in governing variation in speech production.

The Stuff of Change: General Extenders in Toronto, Canada

This article examines general extenders (GEs) in the English spoken in Toronto, Canada, using a 1.2-million-word corpus stratified by age, sex, and education. Employing quantitative techniques, the authors assess the nature of the system, particularly the possibility that it has undergone recent grammaticalization. Diagnostic tests for phonetic reduction, decategorization, semantic change, and pragmatic shift reveal that only decategorization is visible in apparent time. Otherwise, older and younger speakers share most of the same patterns. Yet there is a dramatic shift happening in that the form stuff is rapidly becoming the predominant GE.The authors conclude that in contrast to the United Kingdom, the GEs in Toronto are not grammaticalizing but are undergoing lexical replacement. These findings suggest that discourse-pragmatic features may differ markedly across varieties and further that putative indicators of grammaticalization may not always operate in tandem. GEs provide a unique opportunity to study social and linguistic influences on discourse-pragmatic variation.

Sign lowering and phonetic reduction in American Sign Language

This study examines sign lowering as a form of phonetic reduction in American Sign Language. Phonetic reduction occurs in the course of normal language production, when instead of producing a carefully articulated form of a word, the language user produces a less clearly articulated form. When signs are produced in context by native signers, they often differ from the citation forms of signs. In some cases, phonetic reduction is manifested as a sign being produced at a lower location than in the citation form. Sign lowering has been documented previously, but this is the first study to examine it in phonetic detail. The data presented here are tokens of the sign WONDER, as produced by six native signers, in two phonetic contexts and at three signing rates, which were captured by optoelectronic motion capture. The results indicate that sign lowering occurred for all signers, according to the factors we manipulated. Sign production was affected by several phonetic factors that also influence speech production, namely, production rate, phonetic context, and position within an utterance. In addition, we have discovered interesting variations in sign production, which could underlie distinctions in signing style, analogous to accent or voice quality in speech.

Lexical frequency and Japanese vowel devoicing.

The current study aims to examine the relationship between Japanese vowel devoicing and lexical frequency in light of the probabilistic reduction hypothesis [Jurafsky et al. (2001)], which predicts that more probable words are (phonetically) reduced more than less probable words. In Tokyo Japanese, high vowels become devoiced when they occur between two voiceless consonants. Similar devoicing processes occur in many languages and are generally considered to be part of the vowel reduction processes [e.g., Dauer (1980); Kohler (1990)]. Although Japanese vowel devoicing has been studied extensively from both phonetic and phonological viewpoints, its lexical property, namely, the effect of lexical frequency, is as of yet unknown. If Japanese devoicing is a phonetic reduction process, we would expect to see a stronger degree of devoicing among words with higher lexical frequency. Twenty-four speakers of Tokyo Japanese recorded 60 words which include various devoicing environments. Contrary to the prediction by the probabilistic reduction hypothesis, our preliminary results revealed a significantly higher rate of devoicing [F(1,22)=24.89, p<0.01] as well as shorter vowel duration [F(1,22)=4.61, p<0.05] among low-frequency words than high-frequency words. The implications of these results on the nature of Japanese vowel devoicing will be discussed.

Degrammaticalization

Reviewed by: Degrammaticalization Paolo Ramat Degrammaticalization. By Muriel Norde. Oxford: Oxford University Press, 2009. Pp. xviii, 270. ISBN 9780199207930. $39.95. Before this comprehensive book, Muriel Norde had already published an extensive number of articles on grammaticalization and related phenomena. She now offers in the introductory chapter of this book an exhaustive report on the state of the art concerning grammaticalization and its (alleged) counterpart degrammaticalization. The references extend from p. 239 to p. 258 and no relevant contribution is neglected. The introduction (Ch. 1) contains preliminary definitions of grammaticalization and degrammaticalization, and discusses some related concepts such as lexicalization, reanalysis, and analogy, and controversial types of grammaticalization, such as ‘secondary’ grammaticalization, pragmaticalization, and clause combining. After a short section dedicated to context and constructions, the final part of the introduction deals with methodological issues: is grammaticalization a theory? N agrees with Bernd Heine, Tania Kuteva, and Östen Dahl that grammaticalization can be ‘regarded as a process with a dynamics of its own’ (31). This position is contrary to Newmeyer’s (1998) and Fischer’s (2000) view that grammaticalization is ‘essentially an epiphenomenal result of independent historical developments’ (Newmeyer 1998:235) such as semantic change, phonetic reduction, reanalysis, and metonymic and metaphorical processes. In Fischer’s words, ‘I still cannot see that there is room for a separate or “independent” process of grammaticalization’ (2000:152). In Ch. 2, ‘Unidirectionality’, N directly tackles the problem of degrammaticalization and argues that at no linguistic level can unidirectionality (which means irreversibility of grammaticalization paths) be stated in an absolute way. The aim of the book is to prove that degrammaticalization (e.g. from a grammatical word to a content item) does exist at all language levels. In a previous article of mine against grammaticalization as a unidirectional path (Ramat 2001), I gave a fairly long list of examples of degrammaticalization and commented that ‘[t]he list of degrammaticalizations could be continued, revealing that this process, although not as frequent as its reverse, is by no means irrelevant in language and therefore it cannot be simply dismissed as “exception to the general rule of grammaticalization” ’ (396). Other counterexamples to unidirectionality can be drawn from the perspective of discourse markers as illustrated by Schiffrin (1987). For instance, English deed of indeed was originally a lexical noun and meant ‘action’. In Early Middle English (EMdE) it is often used as a formulaic expression (‘in action’) in PPs. These contexts allow us to think that an event ‘in action’ has to be observable, too. This inference entails the emergence of an epistemic meaning: ofte in storial mateer scripture rehersith the commune opynyoun of men, and affirmeth not, that it was so in dede (lit. ‘often in historical matters scripture repeats the common opinion of men, and affirms not, that it was so in fact’) ‘often where matters of history are concerned, scripture repeats men’s common opinion, but does not affirm that it was so in actuality [rather than opinion]’ (from Traugott 1995:8; source: c. 1388 Purvey Wycliffe Prol I, 56 (HC)). At its first stage, indeed is an adverbial phrase, frequently used in contrast to ‘opinion’ or ‘untruth’. The syntactic sentential scope is reached at the beginning of the EMdE period when indeed reinforces another adversative conjunction in sentence position: somtyme purposely suffring [allowing] the more noble children to vainquysshe, and, as it were, gyuying to them place and soueraintie, thoughe in dede the inferiour chyldren haue more lernyng (Traugott 1995:9; source: 1531 Governor, p. 21 (HC)). [End Page 979] By the seventeenth century indeed starts appearing alone in clause-initial position, that is, in discursive—and not in grammatical—function. It serves to confirm or to amplify a previous statement. Moreover, the inferential process that this marker undergoes is closely dependent not only on the speaker evaluation, but also on the hearer’s recognition (subjectification and intersubjectification). A quotation from Harris and Campbell (1995), which I used in Ramat 2001, is of methodological relevance: ‘no reasonable theory can ignore data just because they are inconvenient: an adequate theory must account for infrequent phenomena, not merely for the most common patterns’ (338). Similarly, N critiques (106f.) Haspelmath, who claims: ‘If one is interested in...

Sensitivity to grammatical and sociophonetic variability in perception

AbstractPhonetic realizations vary depending on social characteristics of the speaker, and recent research provides evidence that individuals are sensitive to at least some of these sociophonetic relationships during perception (Strand and Johnson, Gradient and visual speaker normalization in the perception of fricatives, Mouton, 1996; Hay et al., Journal of Phonetics 34: 458–484, 2006). In addition to socially-conditioned variation, there is evidence that phonetic realizations in production vary depending on the grammatical function of a word (Plug, Phonetic reduction and categorisation in exemplar-based representation:Observations on a Dutch discourse marker, 2005; Hay and Bresnan, The Linguistic Review 23: 321–349, 2006), yet it is not known whether listeners can actively exploit this phonetic variation in speech perception. This paper reports on three perception experiments conducted to determine whether perceivers' sensitivity to fine phonetic detail can assist in extraction of both grammatical and social meaning from the signal.

Two-stage implementation for phonetic recognition using a bi-directional target-filtering model of speech coarticulation and reduction

A structured generative model of a speech coarticulation and reduction is described with a novel two-stage implementation. At the first stage, the dynamics of formants or vocal tract resonance (VTR) are generated using prior information of resonance targets in the phone sequence. Bi-directional temporal filtering with finite impulse response (FIR) is applied to the segmental target sequence as the FIR filter's input. At the second stage the dynamics of speech cepstra are predicted analytically based on the FIR filtered VTR targets. The combined system of these two stages thus generates correlated and causally related VTR and cepstral dynamics where phonetic reduction is represented explicitly in the hidden resonance space and implicitly in the observed cepstral space. The combined system also gives the acoustic observation probability given a phone sequence. Using this probability, different phone sequences can be compared and ranked in terms of their respective probability values. This then permits the use of the model for phonetic recognition.

Spoken Corpora and Analysis of Natural Speech

This paper introduces spoken corpora of Taiwan Mandarin created at Academia Sinica and gives an overview of some recent studies carried out utilizing the spoken data. Spoken language resources of Taiwan Mandarin have been collected and processed at Academia Sinica since 2001. As a result, spoken data, which are useful not only for language archives purpose, but also for linguistic studies, has been made available. In addition to creation of the corpus, two lines of research are discussed in which theoretical and empirical studies are connected by using the aforementioned language resources: 1) language variation and change and 2) spoken discourse analysis. Phonetic reduction is one of the main reasons for changes within a language and it is important to take into account different levels of variations in spontaneous speech. For this purpose, we studied syllable contraction/merger, vowel reduction, and phonetic reduction in directional complements. Discourse items also play an essential part, because they add specific implications to sentences and their use is mainly marked by prosodic means. We segmented a spoken discourse into smaller prosodic units to allow for a more precise study of discourse items, prosodic features, and disfluency. These issues are correlated with each other, especially through prosodic markings.

Second mention reduction in Indian, English, and Korean

Phonetic reduction of the second relative to the first mention of a word in a discourse (Second Mention Reduction, or SMR) is a well-documented feature of spoken American English (AE). This phenomenon could be mediated by discourse-level prosodic structure: second mentions are reduced because they are less likely than first mentions to be accented in AE; or it could occur independently of prosodic structure. Impressionistic studies indicate that Indian English (IE) speakers do not deaccent second mentions like AE speakers [J. Gumperz, Discourse Strategies (1982)]. Korean marks information structure through phrasing and word order rather than pitch accents. Therefore, SMR in these languages would support a direct link between lexical probability and duration. Six IE speakers read five paragraphs containing 59 repeated words, and five Korean speakers read a Korean translation of one of these paragraphs. IE speakers were significantly less likely to deaccent second mentions than AE speakers (sign-test, p<.05), and four IE speakers had identical accenting rates for both mentions, providing empirical evidence that IE speakers do not deaccent second mentions. Moreover, significant SMR was found in both Korean (W=43, p<.05) and IE (W=1340, p<.001), suggesting that probabilistic effects on word duration are not mediated by prosodic prominence.

Discourse variation, grammaticalisation and stuff like that1

General extenders (such as and stuff) are analysed here in the speech of adolescents from three English towns. There were no consistent patterns of gender or social class variation in their use, but a clear social class difference in the use of certain forms, with and that favoured by the working‐class speakers and and stuff and and things preferred by the middle‐class adolescents. The most frequent forms were analysed in terms of phonetic reduction, decategorisation, semantic change and pragmatic shift, changes that together make up the process of grammaticalisation. And that and and everything were the most grammaticalised, followed by or something, with and stuff and and things lagging behind. The multifunctionality of the general extenders caused problems for a rigorous analysis of their pragmatic functions. The paper argues that we must consider their functions within the local contexts in which they occur, to take account of their interaction with other linguistic forms. It is also important to avoid generalising about their functions and, instead, to prioritise the fact that as pragmatic particles they are multifunctional. In this data the general extenders had functions in every communicative domain, often simultaneously. The implications for the quantitative analysis of discourse forms are also considered.

Structured speech modeling

Modeling dynamic structure of speech is a novel paradigm in speech recognition research within the generative modeling framework, and it offers a potential to overcome limitations of the current hidden Markov modeling approach. Analogous to structured language models where syntactic structure is exploited to represent long-distance relationships among words , the structured speech model described in this paper makes use of the dynamic structure in the hidden vocal tract resonance space to characterize long-span contextual influence among phonetic units. A general overview is provided first on hierarchically classified types of dynamic speech models in the literature. A detailed account is then given for a specific model type called the hidden trajectory model, and we describe detailed steps of model construction and the parameter estimation algorithms. We show how the use of resonance target parameters and their temporal filtering enables joint modeling of long-span coarticulation and phonetic reduction effects. Experiments on phonetic recognition evaluation demonstrate superior recognizer performance over a modern hidden Markov model-based system. Error analysis shows that the greatest performance gain occurs within the sonorant speech class.

A bidirectional target-filtering model of speech coarticulation and reduction: two-stage implementation for phonetic recognition

A structured generative model of speech coarticulation and reduction is described with a novel two-stage implementation. At the first stage, the dynamics of formants or vocal tract resonances (VTRs) in fluent speech is generated using prior information of resonance targets in the phone sequence, in absence of acoustic data. Bidirectional temporal filtering with finite-impulse response (FIR) is applied to the segmental target sequence as the FIR filter's input, where forward filtering produces anticipatory coarticulation and backward filtering produces regressive coarticulation. The filtering process is shown also to result in realistic resonance-frequency undershooting or reduction for fast-rate and low-effort speech in a contextually assimilated manner. At the second stage, the dynamics of speech cepstra are predicted analytically based on the FIR-filtered and speaker-adapted VTR targets, and the prediction residuals are modeled by Gaussian random variables with trainable parameters. The combined system of these two stages, thus, generates correlated and causally related VTR and cepstral dynamics, where phonetic reduction is represented explicitly in the hidden resonance space and implicitly in the observed cepstral space. We present details of model simulation demonstrating quantitative effects of speaking rate and segment duration on the magnitude of reduction, agreeing closely with experimental measurement results in the acoustic-phonetic literature. This two-stage model is implemented and applied to the TIMIT phonetic recognition task. Using the N-best (N=2000) rescoring paradigm, the new model, which contains only context-independent parameters, is shown to significantly reduce the phone error rate of a standard hidden Markov model (HMM) system under the same experimental conditions.

The Role of Phonological Predictability in Sound Change: Privileged Reduction in Oceanic Reduplicated Substrings

Phonetic reduction is more likely when a word is predictable or recoverable independent of acoustic information. Attributing higher rates of phonetic reduction to lexical predictability has implications for subword domains. A range of historical developments in Oceanic support this position. In reduplication, where the content of the reduplicated substring is wholly predictable and recoverable from the base, the reduplicant undergoes leniting sound changes more readily than the base, and more readily than other prosodically comparable domains. Synchronic consequences of these developments challenge models associating reduplication with the emergence of unmarked structures.

Negation Markers in Tangxi (A Wu Dialect in Chinese)

This paper first introduces the main negatives of Tangxi dialect in Wu dialects. Based on this, it compares the negatives of Tangxi dialect with those of some dialects in Mandarin, Wu, Hui, Xiang dialects as well as Pinghua from phonetic, morphemic, and semantic angles. Finally, it discusses the origins and functions of sentence-final negatives. Some conclusions made in the paper follow as: phonetic reduction is one means of grammaticalization of negatives; negative concepts of bu (_??_), wei (_??_), wu (_??_) fall into three categories: trichotomy, diachotomy, and monochotomy in Chinese dialects; it is common that negative verb wu (_??_) extends to negative adverb wei (_??_) in Chinese dialects; sentence-final negatives usually result from positive-negative questions' omitting constituents after negatives; sentence-final negatives tend to change into interrogative modal particles.

Secant lines in a preliminary study of phonetic reduction of / ɪ / and /ε /

Two subjects from the X-Ray Microbeam Speech Production Database were examined in their production of the vowels /I/ and /epsilon/ in alveolar and dental consonant contexts. Secant lines, or first-order splines, between the three most anterior pellets were examined at vowel critical times. These critical times were zero crossings in the tangential acceleration of the midpoints of the secant lines. We expected and found, in general, that vowel reduction occurred as a function of vowel duration in measures of the secant line midpoint-to-palate distance and secant line orientation at vowel critical times. The shorter the vowel, the smaller the distance of the secant line midpoints to the palate and the less downward the orientation of the secant lines at the vowel critical times. Phonetic reduction was also apparent in the formant frequencies. There were differences between the speakers in terms of the range of vowel duration and degree of reduction. The subjects differed in the functional parts of the tongue spanned by the secant lines and the shape of their palates. These differences were factors in the observed relations between formant frequencies and the articulatory, secant line measures for each subject.

Essays in Algonquian, Catawban, and Siouan Linguistics in Memory of Frank T. Siebert, Jr.

Blair A. Rudes and David J. Costa (eds.), Algonquian and Iroquoian Linguistics Memoir 16, Winnipeg: Algonquian and Iroquoian Linguistics, 2003, ix + 296 pages.Reviewer: Paul ProulxAs a specialist in Algonquian, I will limit my comments to Algonquian papers. Three are written by scholars generally found lunching together at Algonquian conferences: Ives Goddard, group's leader; Richard Rhodes, his disciple; David Costa, a student of Rhodes and Algonquianist co-editor of volume. The signs of intellectual inbreeding are everywhere, and Goddard clearly suffered from too gentle an editorial hand.Group isolation and status within it can be estimated from volume bibliography, which indexes frequency of citation. 1 count 31 separate Goddard publications cited, versus only seven for Bloomfield. Rhodes has 10 and Costa six, while Peter Denny, a psychologist (these days converted into an amateur archaeologist) who hangs out with them, and remaining linguist, Philip LeSourd, are tied at four.Goddard's contribution is of monograph length (65 pages). Its merit lies in assembling a huge amount of data on Algonquian demonstratives. However, rather than comparing all Algonquian languages, as comparative method calls for, Goddard bases his reconstruction of Proto Algonquian (PA hereafter) demonstratives only on his two fieldwork languages (and some neighbouring ones he regards as closely related to them).These he reconstructs internally, producing something like a 1960s underlying structure. This resemble an earlier stage of a language, or not. In any event, it is a synchronic representation, hypothesizes in simplified and abstract terms about daughter studied, and how to get from there to its more complex concrete (surface) structure.Goddard then reconstructs comparing two underlying systems, and argues away all rest of data in second half of paper. Once this choice of sample and method are made, nothing like a genuine reconstruction is possible. He has discarded far too much information.For example, Goddard claims demonstratives may undergo phonetic reduction beyond what can be accounted for by general sound laws of a language (p. 80). He proposes ad hoc deletion of whole demonstrative roots (VC-) in some languages. In Ojibwa, PA *4eyo(:)- (set A) was reduced to CO *4o-, and *4en- (set B) was completely (p. 63); in Cree-Montagnais, word-initially *4ey- was lost (p. 69); in Menominee, the sequences *4eyo:- and *4eni- were reduced to M *4a- (p. 76); and in Miami, initial I 4iy- is in set A, and I 4n- is in set B (p. 78).I submit this is not just wrong, it's utter nonsense. Even if one believes in ad hoc sound change, wherever Goddard claims reduction has totally deleted a root, he has no evidence root was ever there in first place. Arguably, it is simply a convenient fiction, to avoid recognizing following element is itself a root, and thus existence of additional stems. (For those who believe in regularity of sound change, where his ad hoc reduction has totally deleted a root, although regular sound change would not have done so, he has proven it was never there.)Many of his other claims also go against linguistic intuition and common sense. For example, he says Fox 4i:niya that (animate, inaccessible) and Eastern Cree 4(a)niya: that (animate, inaccessible) do not attest a stem *4eniy- as suggested by Pentland, but rather are separate parallel compoundings of his roots *4en- and *4ey- (p. 38, 71, 91).However, notice attested stems are not Fox *4i:n i:ya nor Eastern Cree *4(a)n-aya:, as a word-initial position for second syllable would have produced by his own rules. The compounding, if such it was, clearly came early, and one wonders by what criteria he refuses to reconstruct it for his Western dialect of (in which *4e > *4i). …

Efficient decoding strategies for conversational speech recognition using a constrained nonlinear state-space model

In this paper, we present two efficient strategies for likelihood computation and decoding in a continuous speech recognizer using an underlying nonlinear state-space dynamic model for the hidden speech dynamics. The state-space model has been specially constructed so as to be suitable for the conversational or casual style of speech where phonetic reduction abounds. Two specific decoding algorithms, based on optimal state-sequence estimation for the nonlinear state-space model, are derived, implemented, and evaluated. They successfully overcome the exponential growth in the original search paths by using the path-merging approaches derived from Bayes' rule. We have tested and compared the two algorithms using the speech data from the Switchboard corpus, confirming their effectiveness. Conversational speech recognition experiments using the Switchboard corpus further demonstrated that the use of the new decoding strategies is capable of reducing the recognizer's word error rate compared with two baseline recognizers, including the HMM system and the nonlinear state-space model using the HMM-produced phonetic boundaries, under identical test conditions.

Context, Culture, and Structuration in the Languages of Australia

▪ Abstract Using Australian languages as examples, cultural selection is shown to shape linguistic structure through invisible hand processes that pattern the unintended outcomes (structures in the system of shared linguistic norms) of intentional actions (particular utterances by individual agents).Examples of the emergence of culturally patterned structure through use are drawn from various levels: the semantics of the lexicon, grammaticalized kin-related categories, and culture-specific organizations of sociolinguistic diversity, such as moiety lects, “mother-in-law” registers, and triangular kin terms. These phenomena result from a complex of diachronic processes that adapt linguistic structures to culture-specific concepts and practices, such as ritualization and phonetic reduction of frequently used sequences, the input of shared cultural knowledge into pragmatic interpretation, semanticization of originally context-dependent inferences, and the input of linguistic ideologies into the systematization of lectal variants. Some of these processes, such as the emergence of subsection terminology and moiety lects, operate over speech communities that transcend any single language and can only be explained if the relevant processes take the multilingual speech community as their domain of operation.Taken together, the cases considered here provide strong evidence against nativist assumptions that see linguistic structures simply as instantiations of biologically given “mentalese” concepts already present in the mind of every child and give evidence in favor of a view that sees individual language structures as also conditioned by historical processes, of which functional adaptation of various kinds is most important. They also illustrate how, in the domain of language, stable socially shared structures can emerge from the summed effects of many communicative micro-events by individual agents.

A path-stack algorithm for optimizing dynamic regimes in a statistical hidden dynamic model of speech

In this paper we report our recent research whose goal is to improve the performance of a novel speech recognizer based on an underlying statistical hidden dynamic model of phonetic reduction in the production of conversational speech. We have developed a path-stack search algorithm which efficiently computes the likelihood of any observation utterance while optimizing the dynamic regimes in the speech model. The effectiveness of the algorithm is tested on the speech data in the Switchboard corpus, in which the optimized dynamic regimes computed from the algorithm are compared with those from exhaustive search. We also present speech recognition results on the Switchboard corpus that demonstrate improvements of the recognizer’s performance compared with the use of the dynamic regimes heuristically set from the phone segmentation by a state-of-the-art hidden Markov model (HMM) system.