Please use this identifier to cite or link to this item:
|Title:||Nonlinear Adaptive Prediction of Nonstationary Signals and its Application to Speech Communications|
|Department:||Electrical and Computer Engineering|
|Keywords:||Electrical and Computer Engineering;Electrical and Computer Engineering|
|Abstract:||<p>Prediction of a signal is synonymous with modeling of the underlying physical mechanism responsible for its generation. Many of the physical signals encountered in practice exhibit two distinct characteristics: nonlinearity and nonstationary. Consider, for example, the important case of speech signals. The production of a speech signal is known to be the result of a dynamic process that is both nonlinear and nonstationary. To deal with the nonstationary nature of signals, the customary practice is to invoke the use of adaptive filtering. Unfortunately, the nonlinear nature of the signal generation process has not received the attention which it deserves in that much of the literature on the prediction of speech signals has focused almost exclusively on the use of linear adaptive filtering schemes.</p> <p>This thesis is aimed at the study of nonlinear adaptive prediction of nonstationary signals using neural networks and its application to real-time speech communication. In this thesis, three basic questions are answered: 1) What kind of neutral networks are suited for real-time adaptive signal processing? 2) How can an adaptive neural network predictor be designed? 3) Can a neural network predictor be used for the application of real-time communication?</p> <p>In this thesis, a new Pipelined Recurrent Neural Network (PRNN) is designed. The PRNN is composed of M separate modules. The modules are identical, each designed as a recurrent neural network with a single output neuron. Information flow into and out of the modules proceeds in a synchronized fashion.</p> <p>A new scheme for the nonlinear adaptive prediction of nonstationary signals is proposed. The nonlinear neural network-based filter, which consists of a PRNN and a linear filter, learns to adapt to statistical variations of the incoming time series while, at the same time, the prediction is going on. The dynamic behavior of the pipelined recurrent neural network-based predictor is demonstrated in the case of several speech signals; for these applications it is shown that the nonlinear adaptive predictor outperforms the traditional linear adaptive scheme in a significant way. It should however, be emphasized that the nonlinear adaptive predictor has a much wider range of applications such as the adaptive prediction of sea clutter.</p> <p>The PRNN-based adaptive predictor is applied to the adaptive differential pulse code modulation. In the encoder and decoder parts of an ADPCM system, the predictor is successfully incorporated with an adaptive quantizer for low bit-rate speech communication. The research work involves a novel combination of pipelined recurrent neural network and a robust linear adaptive filter, and the design of a new 4-, 8- or 16-level adaptive quantizer. The nonlinear adaptive differential pulse code modulation algorithm is tested with different speech signals. The new algorithm is compared with a linear ADPCM algorithm, recommended by CCITT, from several aspects such as time domain, frequency domain, and listening tests. Speech experiments show that nonlinear adaptive differential pulse code modulation provides a promising new approach for high-quality communication at low bits rates.</p>|
|Appears in Collections:||Open Access Dissertations and Theses|
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.