Acoustic target zones for naturally produced vowels in running speech
Frank E. Kramer, Donald J. Meyer, Marlos Fourskis, James D. Miller · The Journal of the Acoustical Society of America · 1991
Acoustic measures of 2304 vowel tokens of Midwestern American English were collected by one of us (Fourakis). The vowels [i, ɪ, ɛ, æ, ɑ, ɔ ᴜ, u, ʌ] occurred in [bVd] or [hVd] syllables imbedded in sentences spoken with different rates of speech and different stress patterns. There were four male and four female speakers. Two methods of automatic classification of this corpus of acoustic measures into vowel categories are evaluated. One method derives from parametric statistical theory and has been used often in studies of speech sounds. This method is linear discriminant analysis (LDA). The other method assigns each point in the acoustic space a most likely “vowel classification” by a “spatial windowing” procedure. Then an edge-detection algorithm automatically finds the edges of regions with similarly classified interior points. This is called the target-zone method (TZM). Both methods are applied completely automatically to the acoustic descriptions of the vowels. When the vowels are described acoustically in terms of Miller's auditory-perceptual space [J. D. Miller, J. Acoust. Soc. Am. 85, 2114-2134 (1989)], the target zones automatically generated by TZM classify about 85% of the tokens correctly. The LDA method does not achieve this level of performance.