Acoustic signal processing based on the short-time spectrum

The frequency domain representation of a time signal afforded by the Fourier transform is a powerful tool in acoustic signal processing. The usefulness of this representation is rooted in the mechanisms of sound production and perception. Many sources of sound exhibit normal modes or natural frequencies of vibration, and can be described concisely in the frequency domain. The human auditory system performs frequency analysis early in the hearing process, so perception is often best described by frequency domain parameters. This dissertation investigates a new approach to acoustic signal processing based on the short-time fourier transform, a two dimensional representation which shows the time and frequency structure of sounds. This representation is appropriate for signals such as speech and music. Where the natural frequencies of the source change and timing of these changes is important to perception. The principal advantage of this approach is that the signal processing domain is similar to the perceptual domain, so that signal modifications can be related to perceptual criteria. The mathematical basis for this type of processing is developed, and four examples are described: removal of broad band background noise, isolation of perceptually important speech features, dynamic range compression and expansion, and removal of locally periodic interfering signals.