Perceptual Cepstral filters for speech and music processing

Source-filter modeling of speech or musical tones requires a filter model for the spectral envelope of the signal. To reduce the number of modeling parameters, one idea is the use of psychoacoustic knowledge to encode only the relevant information in a perceptual sense. Starting from an accurate estimation of the original spectral envelope, with imperceptible details, in this work, we propose to use its Mel-Frequency Cepstral Coefficient (MFCC) representation to catch the perceptually relevant information. Then, a new inverse process is presented to derive a smoother, but perceptually equivalent spectral envelope. For instance, this new method can be applied in speech coding, and thanks to the good properties of the MFCC representation, perceptual interpolations of sounds is made easier.