<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1046">
  <Title>Word Boundary Identification fro m Phoneme Sequence Constraints in Automatic Continuous Speech Recognition</Title>
  <Section position="1" start_page="0" end_page="0" type="abstr">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> Th.s paper explores the extent to which phoneme sequence constralnt'~ can be used to identify word boundaries in coutinnous speech recog~fition. The input consists of phonemic transcriptions (without word boundaries indicated) of 145 utterances produced by 1 ~e speaker. The constraints are derived by matching the complete set of 3 phonente sequences that can occur' across word boundaries to entries in large lexicons containing both citation and reduced h)rm pronunciations. Phonemic assimilatoxy adjushnents across word boundaries are alse taken into account. The results show that around 37% of all word boundaries can be correctly identified t rom a knowledge of .~uch phoneme sequence contraints alone, and ~hat this figure rises to 45% when a knowledge of oneand two-phoneme words and all legal, word-initial and word-final, two-pitoneloe sequences are taken into account. The possibility of including :~uch constraints in tim architecture of a corrtinuous speech reeogniser is discussed.</Paragraph>
    <Paragraph position="1"> I. Introduction TiLe identification of word boundaries fl'om continuous speech by human listeners depends, in part, on an interaction between :prosodic, syntactic and semantic processing, Since, however, ~his interaction is difficult to model in machines and since some prosodic variables, such as sentence stress patterns, are difficult to extract automatically from the acoustic waveform, the identification of word bmmdaries must often be accomplished by different hinds of processing in continuous ,~peeciL reeognisers: one possibility~ discussed in Lamel &amp; Zue (1984) and explored in this paper, depends on the incorporation of a knowledge of' phoneme sequence constraints, Phoneme sequence constraints are based on a knowledge of phoneme sequences which do not occur ! word-internally: for' example, since there are no words which end in/m g/~ at~d since/m g l/does not occur word-internally, a word boundary must occur after/m/ (Lamel &amp; Zue, 1984). Harrington, Johnson &amp; Cooper (1987) showed that word boundary CVC sequences are often excluded word-internally in monomorphemic words if the pre- and post-vocalic consonants are similar: thus,/s N V N/(N == nasal),/C l V l/,/f V p/,/g V ld,/z V ,iJ,/sh V sh/ are all exchtded, or are at least extremely rare, word-internally in British English Received Pronunciation (top). In the study discussed b.~low, we extend the investigations of Lamel &amp; Zue i (1984) anti 1 larrington et al. (1987) by developing an algorithm for the autmmttic identification o~&amp;quot; word boundaries from such sequences in a continuous speech recogniser.</Paragraph>
    <Paragraph position="2"> In the Alvey Demonstrator continuous speech recogniser being developed at the Centre for Speech Technology Research (CSTR), Edinburgh University (Figure 1), the identification of word boundaries from a string of phonemes is accomplished by a chart-parsing s~rategy which matches the lexicon from left-to-right against a string of phonemic symbols that are themselve~ derived from the phonetic processing of the acoustic-waveform. In this system, only cmnplete parsings of the phonemic units are passed to higher' levels for syntactic and semantic processing. The only possible parsing, therefore, of the phonemic string/t ii eh i ng w i 1/is teaching+will, since there are no ether paths which parse the entire string of phonemes.</Paragraph>
    <Paragraph position="3">  The relationship between the identification of word boundaries fl'om piloneme sequence constraints and the chart-parsing strategy outlined above can be clarified with respect to Figure 1: at all points where the arcs do not overlap, it should be possible to in.~ert a word bour~dary from a knowledge of ptloneme sequence constraints. Since, therefore, the only point at which the arcs are non-overlapping is between /ng/ and /w/, phoneme sequence constraints stlould apply to insert a word boundary at that point (there being no monomorphemic words in the English language that contain a medial /ng w/). At the same time, however, Figure 1 would seem to suggest that the prior&amp;quot; implementation of phoneme sequence constraints is superfluous, since all word boundaries can he found frmn the chart-parsing strategy. However, the application of phoneme sequence constraints may enable recovery when the chart-parsing strategy is unable to parse the phonemic string because of the incorrect :derivation of a particular phoneme. Suppose, for example, that the acoustic-phonetic component incorrectly derives/el ng/from the parameteriscd acoustic wavetbrm instead of/i ng/(Figure 2).</Paragraph>
  </Section>
class="xml-element"></Paper>