<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-1094">
  <Title>A concurrent approach to the automatic extraction of subsegmental primes and phonological constituents from speech</Title>
  <Section position="6" start_page="581" end_page="581" type="concl">
    <SectionTitle>
4 PhonMaster and its successors
</SectionTitle>
    <Paragraph position="0"> The PhonMaster prototype was implemented in C++ by a PhD student educated in object-oriented design and Windows application programming. It uses standard object-class libraries for screen management, standard relational database tools for control of the lexicon and standard code for FFT as in a spectrogram display object. Users may add words using a keypad labelled with IPA symbols.</Paragraph>
    <Paragraph position="1"> Manner class sequences and constituent structure are generated automatically. The objects concerned wilh the extraction of cues from spectra, segmentation, manner-class sequencing and display of constituent structure, repairing effects of lenition and assimilation are custom built.</Paragraph>
    <Paragraph position="2"> PhonMaster does not use corpus trigram statistics (e.g. Young 1996) to disambiguate word lattices, and there is no speaker-adaptation. Without these standard ways of enhancing pure pattern-recognition accuracy, its success rate for pure word recognition is around 75%. We are contemplating the addition d&amp;quot; pitch cues, which, with duration, would allow detection of stress, which may further increase accuracy.</Paragraph>
    <Paragraph position="3"> Object orientation makes the task of incorporating currently popular pattern recognition methods fairly straightforward. HMMs whose hidden states have cues like ours as observables are obvious things to try. Artificial Neural Nets (ANNs) also fit into the task architecture in various places. Vector quantisation ANNs could be used to learn the best choice of thresholds for head-operator detection and discrimination. ANNs with output nodes based on our quadratic discriminants in place of the more common linear discriminants are also an option, and their output node strengths would be direct measures of presence of elements.</Paragraph>
  </Section>
class="xml-element"></Paper>