| Conventional speech analysis techniques are based on estimating the spectral content of relatively short (about 10--20 ins) segments of the signal. However, an alternate way to describe a speech signal is a summation of amplitude modulated frequency bands, where each frequency band consists of a smooth envelope (gross structure) modulating a carrier signal (fine structure). The analytic signal (AS) forms a suitable candidate for such an envelope-carrier decomposition with the squared magnitude of the AS, called the Hilbert envelope, representing the smooth structure and the phase component of the AS representing the fine structure. However, the computation of analytic signal is cumbersome and theoretically requires the use of a filter with infinite impulse response.;In this thesis, we adopt an auto-regressive (AR) modeling approach for estimating the Hilbert envelope of the signal. The Hilbert envelope represents the evolution of signal energy in time domain. This model, referred to as frequency domain linear prediction (FDLP), is based on the application of linear prediction on discrete cosine transform of the signal. Thus. FDLP is dual process to the conventional time domain linear prediction (TDLP).;Just like conventional AR models, the FDLP model describes the perceptually dominant peaks and removes the finer-scale detail. This suppression of detail is particularly useful for parametric representation of speech/audio signals, where the goal is to summarize the general form of the signal. We show several applications of the FDLP model for speech and audio processing systems. As a unified model of speech and audio signals, we apply the FDLP technique for wide-band high fidelity audio coding. In subjective evaluations, the FDLP codec compares well with state-of-art speech/audio codecs.;In order to derive robust representation in the presence of reverberation and channel distortions, we propose a gain normalization procedure for FDLP envelopes. The gain normalization suppresses the effect of long-term convolutive distortions in sub-bands of speech. We apply the gain-normalized FDLP envelopes for feature extraction in speaker, speech and phoneme recognition experiments. In these experiments, the FDLP features provide significant improvements over the conventional techniques in noisy and reverberant environments. |
