<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-1611">
  <Title>A Transcription Scheme for Languages Employing the Arabic Script Motivated by Speech Processing Application</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Speech-to-speech (S2S) translation systems present many challenges, not only due to the complex nature of the individual technologies involved, but also due to the intricate interaction that these technologies have to achieve. A great challenge for the specific S2S translation system involving Persian and English would arise from not only the linguistics differences between the two languages but also from the limited amount of data available for Persian. The other major hurdle in achieving a S2S system involving these languages is the Persian writing system, which is based on the Arabic script, and hence lacks the explicit inclusion of vowel sounds, resulting in a very large amount of one-to-many mappings from transcription to acoustic and semantic representations.</Paragraph>
    <Paragraph position="1"> In order to achieve our goal, the system that was designed comprised of the following components: Fig 1. Block diagram of the system. Note that the communication server allows interaction between all subsystems and the broadcast of messages. Our vision is that only the doctor will have access to the GUI and the patient will only be given a phone handset.</Paragraph>
    <Paragraph position="2">  (1) a visual and control Graphical User Interface (GUI); (2) an Automatic Speech Recognition (ASR) subsystem, which works both using Fixed  State Grammars (FSG) and Language Models (LM), producing n-best lists/lattices along with the decoding confidence scores; (3) a Dialog Manager (DM), which receives the output of the speech recognition and machine translation units and subsequently re-scores'' the data according to the history of the conversation; (4) a Machine Translation (MT) unit, which works in two modes: Classifier based MT and a fully Stochastic MT; and finally (5) a unit selection based Text To Speech synthesizer (TTS), which provides the spoken output. A functional block diagram is shown in Figure 1.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
1.1 The Language Under Investigation:
Persian
</SectionTitle>
      <Paragraph position="0"> Persian is an Indo-European language with a writing system based on the Arabic script.</Paragraph>
      <Paragraph position="1"> Languages that use this script have posed a problem for automated language processing such as speech recognition and translation systems. For instance, the CSLU Labeling Guide (Lander, http://cslu.cse.ogi.edu/corpora/corpPublications.ht ml) offers orthographic and phonetic transcription systems for a wide variety of languages, from German to Spanish with a Latin-based writing system to languages like Mandarin and Cantonese, which use Chinese characters for writing.</Paragraph>
      <Paragraph position="2"> However, there seems to be no standard transcription system for languages like Arabic, Persian, Dari, Urdu and many others, which use the Arabic script (ibid; Kaye, 1876; Kachru, 1987, among others).</Paragraph>
      <Paragraph position="3"> Because Persian and Arabic are different, Persian has modified the writing system and augmented it in order to accommodate the differences. For instance, four letters were added to the original system in order to capture the sounds available in Persian that Arabic does not have. Also, there are a number of homophonic letters in the Persian writing system, i.e., the same sound corresponding to different orthographic representations. This problem is unique to Persian, since in Arabic different orthographic representations represent different sounds. The other problem that is common in all languages using the Arabic script is the existance of a large number of homographic words, i.e., orthographic representations that have a similar form but different pronunciation. This problem arises due to limited vowel presentation in this writing system.</Paragraph>
      <Paragraph position="4"> Examples of the homophones and homographs are represented in Table 1. The words six and lung are examples of homographs, where the identical (transliterated Arabic) orthographic representations (Column 3) correspond to different pronunciations [SeS] and [SoS] respectively (Column 4). The words hundred and dam are examples of homophones, where the two words have similar pronunciation [sad] (Column 4), despite their different spellings (Column 3).</Paragraph>
      <Paragraph position="5">  and their limitation. Purely orthographic transcription schemes (such as USCPers) fail to distinctly represent homographs while purely phonetic ones (such as USCPron) fail to distinctly represent the homophones.</Paragraph>
      <Paragraph position="6"> The former is the sample of the cases in which there is a many-to-one mapping between orthography and pronunciation, a direct result of the basic characteristic of the Arabic script, viz., little to no representation of the vowels.</Paragraph>
      <Paragraph position="7"> As is evident by the data presented in this table, there are two major sources of problems for any speech-to-speech machine translation. In other words, to employ a system with a direct 1-1 mapping between Arabic orthography and a Latin based transcription system (what we refer to as USCPers in our paper) would be highly ambiguous and insufficient to capture distinct words as required by our speech-to-speech translation system, thus resulting in ambiguity at the text-to-speech output level, and internal confusion in the language modelling and machine translation units. The latter, on the other hand, is a representative of the cases in which the same sequence of sounds would correspond to more than one orthographic representation. Therefore, using a pure phonetic transcription, e.g., USCPron, would be acceptable for the Automatic Speech Recognizer (ASR), but not for the Dialog Manager (DM) or the Machine Translator (MT). The goal of this paper is twofold (i) to provide an ASCII based phonemic transcription system similar to the one used in the International Phonetic Alphabet (IPA), in line of Worldbet (Hieronymus, http://cslu.cse.ogi.edu/corpora/corpPublications.ht ml) and (ii) to argue for an ASCII based hybrid transcription scheme, which provides an easy way to transcribe data in languages that use the Arabic script.</Paragraph>
      <Paragraph position="8"> We will proceed in Section 2 to provide the USCPron ASCII based phonemic transcription system that is similar to the one used by the International Phonetic Alphabet (IPA), in line of Worldbet (ibid). In Section 3, we will present the USCPers orthographic scheme, which has a one-to-one mapping to the Arabic script. In Section 4 we will present and analyze USCPers+, a hybrid system that keeps the orthographic information, while providing the vowels. Section 5 discusses some further issues regarding the lack of data.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>