<?xml version="1.0" standalone="yes"?>
<Paper uid="A83-1032">
  <Title>APPLICATION OF THE LIBERMAN-PRINCE STRESS RULES TO COMPUTER SYNTHESIZED SPEECH</Title>
  <Section position="5" start_page="194" end_page="196" type="concl">
    <SectionTitle>
IV EVALUATION
</SectionTitle>
    <Paragraph position="0"> An experiment was conducted to evaluate stress placemenc using the computerized version of the LP model. A random sample of unique English words and their correct phonetic translations used for the axperlment was selected from the American Heritage Dictionary \[i\]. Five hundred pairs of random numbers were generated; the first number in the pair was a random number between one and the page number of the last page in the dictionary and the second one was a random number between one and sixty. For each pair, the first number was the page on which the random word was to be found and the second number, 2, determined the word to be the ~'th on the page. If ~ was larger than the actual number of words on the page, then n modulo the number of words on the page was used. If the selected word was not polysyllabic, It was rejected. Using this technique, 357 unique random words were selected. Each word was translated into ASCII codes for the VOTRAX according to the phonetic translation in the dictionary. These translations were then given as input to the stress system.</Paragraph>
    <Paragraph position="1"> Because the words in the random sample contain combinatlons of primary, secondary, and tertiary stress, several methods arise for evaluatlng the results (listed in the order of importance): null i) The number of words completely correct, the number of words incorrect, and the number of words which dropped through.</Paragraph>
    <Paragraph position="2"> 2) The number of times primary, secondary, and tertiar 7 stress were each individually predicted correctly regardless of the other two.</Paragraph>
    <Paragraph position="3"> 3) The number of times when secondary or tertiary stress was incorrectly predicted.</Paragraph>
    <Paragraph position="4"> 4) The number of rimes secondary or tertiary stress was predicted but the word did not require it.</Paragraph>
    <Paragraph position="5"> 5) The number of times secondary or tertiary stress was needed but not predicted.</Paragraph>
    <Paragraph position="6"> The figures for the first evaluation are shown in Table 2. The totally correct words are slightly under two thirds of the entire sample. However, when the words with correct stress and the words which fell through are combined, the total is slightly over 70X.</Paragraph>
    <Paragraph position="7">  The results of the second evaluation are shown in Table 3. While primary stress is predicted correctly in 75% of the cases, secondary stress is only 53Z and tertiary stress occurs too infrequently to make any observations. The number in parentheses in Table 3 indicates the total number of the particular stress level required. words of Table 2. The importance of this fact appears when one considers that the stress pattern is partially correct, but is not distortec by incorrect stressing. Therefore even though partial, this stress pattern would be an improvement. If these words are now combined with the totally correct words and those which dropped through, they equal 291 words or 81.51%, i.e. almost 82~ of the words can be stressed totally, partially, or left unchanged.</Paragraph>
    <Paragraph position="8"> TABLE 6. Secondary and tertiary stress which was not predicted.</Paragraph>
    <Paragraph position="9">  placed on the wrong syllable is small but still significant. Again tertiary stress occurrences were too few to make observations.</Paragraph>
    <Paragraph position="10">  With 63.3% of the sample words completely correct, 73.10% of the sample words completely or partially correct, 8.4% unmodified and 18.49% in error, this test has demonstrated that the stress model defined by the stress system and its input rules does work in a substantial percentage of cases.</Paragraph>
    <Paragraph position="11"> Of the 66 words that were incorrectly stressed, most fall into one of four categories. I) Two syllable words where the vowel pattern is -lons -lon~ or +lons +lon~ and the last syllable is stressed. In these cases the stress system incorrectly assigns stress to the first vowel: e.g., transact, mistrust.</Paragraph>
    <Paragraph position="12"> 2) Words in which the ESR or SKR skips over syllables that should be stressed, e.g.</Paragraph>
    <Paragraph position="13"> isodynamic, epox-/, comprehend, remitter, inopportune.</Paragraph>
    <Paragraph position="14"> The results of the fourth test are given in  the sample, this is a relatively small number of erroneous predictions.</Paragraph>
    <Paragraph position="15"> TABLE 5. Stress that should not have been predicted.</Paragraph>
    <Paragraph position="16"> Secondary Tertiary # 3 1 Finally the fifth evaluation leads to Table 6. This table shows the number of times secondary or tertiary stress was required but not predicted. An interpretation of this table suggests that for 35 words which needed both primary and secondary stress, only primary stress was predicted. These words are also included in the incorrectly stressed 3) When in a two syllable word, the word stem vowel is short and the prefix or suffix vowel is long, the long vowel is marked for stress, e.g. fancied.</Paragraph>
    <Paragraph position="17"> 4) The LCPR does not correctly assign nodes ~, ~, values, e.g. contumacy, Kastight.</Paragraph>
    <Paragraph position="18"> Each of these groups is an exception to a larger group whose stress patterns fit the predicted patterns.</Paragraph>
    <Paragraph position="19"> A final question is: How well does this system predict stress in the most common English words? Of the 200 most common, 162 have a single vowel in their phonetic translation and therefore would drop through the system without being modified. Of the 38 remaining words, 33 are correctly stressed by the stress system, leaving 5 incorrectly stressed. However, since these are the most common of words of English, it would seem reasonable to include these words as special rules  in the rule system of the translator and not allow the stress system to operate on them.</Paragraph>
    <Paragraph position="20"> V SUI~t~Y Computer synthesized speech and linguistic theories for predicting stress can interact with one another to mutual benefit. Computer synthesized speech techniques can be used to evaluate the linguistic theory. Just as computers have been used so often to evaluate theories in other disclpllnes, so too can ~hey be used in linguistics. The organizationt speed, accuracy and unblasedness of the computer makes it superior to a person in many respects for Judging a hypothesis.</Paragraph>
    <Paragraph position="21"> On the other hand, the linguistic theories can provide a substantial base on which to build language components of artificially intelligent systems. The intelligibility of computer synthesized speech can be improved with the application of linguistic theories for predicting stress such as that proposed by Liberman and Prince. Evaluations such as that presented in this paper will be of value not only in comparing competing theories but will also be helpful in determ/ning whether the accuracy of a theory's predlctions is acceptable for a particular application and where improvements ,my be made to the theory.</Paragraph>
  </Section>
class="xml-element"></Paper>