<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3201">
  <Title>A Combined Phonetic-Phonological Approach to Estimating Cross- Language Phoneme Similarity in an ASR Environment</Title>
  <Section position="2" start_page="0" end_page="1" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Speech technologists typically use acoustic measurements to determine similarity among acoustic speech models (phone(me) HMMs) and there are a variety of distance metrics available that prove the effectiveness of this method (see Sooful and Botha 2002). Additionally, HMM similarity can be evaluated indirectly through comparison of HMM performances in ASR experiments.</Paragraph>
    <Paragraph position="1"> For acoustic measurements, speech data must be accessible for model training. However, speech data unavailability is a practical concern in that most commercially available speech databases are restricted to widely spoken languages in large business markets. The vast majority of languages have not been exposed to intense data collection and resources for these languages are subsequently either limited or completely unavailable. Hence a knowledge-based phoneme distance metric potentially has great value in acoustic modeling for resource-limited languages in that it can predict cross-language HMM similarity in the absence of target-language speech data.</Paragraph>
    <Paragraph position="2"> Knowledge-based approaches to HMM similarity generally attempt to identify articulatory similarity between phonemes across languages. The typical strategy is subjective and label-based, where two phonemes are judged to be more or less similar depending on their transcription labels (Kohler 1996; Schultz and Waibel 1997, 2000).</Paragraph>
    <Paragraph position="3"> A label-based approach suffers for two obvious reasons. First, phone inventories designed for speech technology applications are predominantly phonemic in orientation. Thus, transcription labels do not transfer with the same phonetic value to other languages, even where international phonetic transcription labels are employed. In a phonemic transcription strategy, transcription labels are gen- null erally restricted to only the most basic symbols, usually unmodified letters of the Roman alphabet (IPA 1999). Second, phoneme transcription labels fail to capture allophony. The best phonetic definition that a phoneme transcription label can offer is the most typical phonetic realization of that phoneme. Not surprisingly, label-based cross-language transfer experiments have produced poor performance results.</Paragraph>
    <Paragraph position="4"> In contrast to the subjective, label-based strategy, researchers in such fields as language reconstruction, dialectometry, and child language development, commonly use automatic feature-based approaches to articulatory similarity between phonemes. In these methods, phonemes are represented by a distinctive feature vector and a phonetic distance or similarity algorithm is used to align phoneme strings between related words (Connolly 1997; Kessler 1995, 2005; Kondrak 2002; Nerbonne and Heeringa 1997; Somers 1998). Significantly, in these approaches, phonological similarity is generally assumed.</Paragraph>
    <Paragraph position="5"> In principle, the feature-based approach to phonetic distance admits more precise specification of phonemes because it supports allophonic variance.</Paragraph>
    <Paragraph position="6"> For example, a standard feature-based approach to allophony representation restricts feature inclusion to only those features relevant to all realizations of the phoneme. Another common approach retains features that are relevant to all allophonic variants, but leaves their values underspecified (Archangeli 1988). However, it is unclear from the literature whether allophony is explicitly addressed in the current feature-based approaches to phoneme similarity. null A strategy for specifying allophony and characterizing phonetic distance between phonemes is only one component in predicting phoneme similarity among diverse languages without acoustic data in an ASR environment. Because HMMs represent phonemes and significant allophones in a language-dependent context, it is necessary to consider the overall constructed target-language HMM system. Thus phonological distance quantities that regulate the priority of source languages for phoneme selection in accordance to their phonological similarity to the target language are also in order.</Paragraph>
    <Paragraph position="7"> In this paper, we describe an automated, combined phonetic-phonological (CPP) approach to estimating phoneme similarity across languages in ASR. Elsewhere, we provide the phonetic and phonological distance algorithms (Liu and Melnar 2005, 2006), though offer little linguistic justification of the approach or evaluation of the experiment results due to space limitations. Here, we focus on explaining the linguistic principles behind the algorithms and analyzing the results.</Paragraph>
    <Paragraph position="8"> The CPP approach is fundamentally based on articulatory phonetic features and is designed to handle allophonic variation. Feature salience and phonetic distance are automatically calculated and phoneme distance is constrained by statistically-derived phonological similarity biases. Unlike other distinctive feature-based approaches to phoneme similarity, phonological distance is not assumed. In testing this approach in cross-language transfer experiments, target-language resources are restricted to lexica and phonology descriptions and do not include speech data.</Paragraph>
    <Paragraph position="9"> In the next section, we describe our feature-based phoneme specification method. In section three, we show how our phoneme specification approach is used in calculating phonetic distance between phonemes. Section four describes two other distance metrics that predict phonological similarity between languages. We explain how the three distance metrics combine to quantify cross-language phoneme distance and select target-language phoneme HMM inventories. In section five, we describe the experiments that we conducted to evaluate our approach to phoneme similarity prediction. Here, the CPP method is compared with an acoustic distance method in context-independent speech recognition. We offer our evaluation and conclusions in section 6.</Paragraph>
  </Section>
class="xml-element"></Paper>