Multiscale fourier descriptors for classifying semivowels in spectrograms

Abstract

Fourier descriptors (FDs) have been used successfully to characterize the boundary of objects in images. It is demonstrated that FDs are appropriate for characterizing objects in speech spectrograms consisting of 40 sounds representing 10 speaker-dependent words containing the English semivowels /w y l r/. With 10 FDs, a 97.5% recognition rate is attained. Different sounds are misclassified by wide- and narrow-band methods, suggesting that multiscaling and FD changes (differences) may be appropriate features. With a FD difference approach, recognition rates equaled or exceeded those obtained with a conventional linear predictive coding (LPC) classifier as well as those obtained with wide- and narrow-band FD methods alone.