Narrow Phonetic Transcription for Development of a Large Vocabulary Isolated Word Recognizer

To build a very large vocabulary (50K) isolated word speech recognizer, speech data from over 200 native speakers of American English was recorded and manually transcribed. This paper explains the transcription method used, the motivation, and preliminary results implemented in the recognizer. The symbol set was expanded to allow for narrow phonetic transcriptions, similar to the level of detail allowed by the International Phonetic Alphabet (IPA). By expanding the symbol set, the following was expected: n more accurate symbolic representation of the acoustic material; n greater consistency between transcribers; n greater flexibility in the use of the phonetic transcriptions during training and testing of the speech recognizer; n greater insight into the acoustic-phonetic variables that affect the performance of the speech recognizer; n over all, to build a better recognizer.