Phoneme Set Design for Speech Recognition of English by Japanese

Abstract

SUMMARYThis paper describes a novel method to improve the performance of second language speech recognition when the mother tongue of users is known. Considering that second language speech usually includes less fluent pronunciation and more frequent pronunciation mistakes, the authors propose using a reduced phoneme set generated by a phonetic decision tree (PDT)-based top-down sequential splitting method instead of the canonical one of the second language. The authors verify the efficacy of the proposed method using second language speech collected with a translation game type dialogue-based English CALL system. Experiments show that a speech recognizer achieved higher recognition accuracy with the reduced phoneme set than with the canonical phoneme set. key words: phonetic decision tree (PDT), Phoneme set, Second language speech recognition

Speech Recognition of English by Japanese Using Lexicon Represented by Multiple Reduced Phoneme Sets

“…5 showed that the most highly improved performances appeared at the lowest and highest proficiency levels. The experimental results of our previous studies showed that a single 28-reduced phoneme set provided better performance in comparison with other reduced ones for speech by L2 speakers on average [8]. 25-phoneme and 32-phoneme sets, which have slightly different characters from the 28-phoneme set, were assigned as the proficiency-dependent reduced phoneme set to L2 speakers of lowest and highest proficiency levels, respectively.…”

“…To carry out the preliminary splitting, we utilized 166 discrimination rules [8] that were designed to categorize each phoneme on the basis of phonetic features such as the manner and position of articulation. A selection of these discrimination rules is shown in Table A· 1 (appendix).…”

“…Our previous research demonstrated that phoneme mismatches resulting in mis-recognition of second language speech can be improved by reducing the number of phonemes [8]. The tendency of mispronunciation depends on the average proficiency level of L2 speakers.…”

“…The baseline is ASR using the canonical phoneme set consisting of 41 phonemes. In order to evaluate the relation between the optimal phoneme set number and second language speakers, we also used various reduced phoneme sets with numbers ranging from 21 to 38 obtained from our previous study [8].…”

“…To improve the recognition accuracy for non-native speech in cases where the mother tongue of the speaker is known, we previously proposed using a reduced phoneme set (RPS) created with a phonetic decision tree based topdown sequential splitting method [8]. We applied this method to the recognition of English utterances spoken by Japanese speakers of various English proficiencies and demonstrated that the reduced phoneme set was more effective than the canonical phoneme set conventionally used in English ASR (refer to Sect.…”