A modified extended Kalman filter approach to demodulation of AM-FM signals and its applications to audio and speech signals

A large body of recent research in signal and image processing has involved mathematical models in which the measured signal is the result of simultaneous amplitude (AM) and frequency (FM) modulation. In this thesis we focus on two aspects of this area: new statistical methodology for describing the AM and FM modulations and extracting the modulating signals from the measured signal and applications of this methodology to speech and audio coding. In the methodology we have developed tools that force the AM modulating signal to be positive and force both the AM and FM modulating signals to have low bandwidth while extracting the AM and FM modulating signals from the measured signals by using statistical extended Kalman filtering (EKF) methods. In the application we have proposed a complete speech coding system based on applying our methodology to the outputs of a fixed filter bank followed by perceptual coding and scalar quantization of the resulting AM and FM modulations. The methods and application are well-coupled because one of the tools for bandwidth control of the AM and FM modulating signals is interpolation over fixed duration intervals and such intervals, under the name of “frames”, have long been an important component of speech analysis. We performed a subjective listening test in which 18 subjects listened to 21 pairs of sentences and compared the quality between our new coding algorithm and a standard coding algorithm