A study of vowel normalization and identification in connected speech

This paper discusses the identification of vowels in connected speech for a recognition system in which no prior knowledge about the input speakers is assumed. Portions of vowels in ten sentence utterances produced by male and female speakers were labeled on the basis of both acoustic analysis and listening. Formant frequencies and vocal tract area functions were extracted for these vowel portions by the LPC method. Also, the vocal tract length for each analysis frame was estimated acoustically and utilized to normalize the above parameters. The effectiveness of the normalization was investigated by conducting identification experiments based on reference data obtained from 26 speakers. Several factors affecting the identification are discussed.