Structures for coding noisy speech at medium-to-low bit rates /
| Main Author: | |
|---|---|
| Other Authors: | , , |
| Format: | Thesis Book |
| Language: | English |
| Published: |
1988.
|
| Subjects: | |
| Online Access: | ProQuest, Abstract Link to OAKTrust copy |
| Abstract: | Encoding of noisy speech at medium-to-low bit rates is considered in this dissertation. Ths cascade estimator and coder scheme is adopted based on the previous result that the overall mean-squared error can be decomposed into an estimator part and a coder part. The vector model Kalman filters are implemented as a consequence of the alphabet-constrained formulation, and the scalar model Kalman filters are implemented as a suboptimal version, for both white noise and colored noise problems. Three types of Kalman filter(KF) parameter estimators are presented. The expectation-maximization(EM) method yielded near-optimal KF performance converging to the ideal counterpart within 0.5 dB in signal-to-noise ratio(SNR). For one-path but suboptimal estimation, the adaptive prediction(AP) method is presented, which performs adequately when its simplicity is considered. Another suboptimal method, applicable only to the white noise problem, yields similar SNR compared to the AP method but somewhat different speech quality. For vector quantization(VQ) and linear predictive coding(LPC) of noisy speech, it is shown experimentally that the near-optimality of the EM method is well-preserved at the final VQ or LPC output. In evaluating performance, SNR's are supplemented by informal listening tests and spectrograms. |
|---|---|
| Item Description: | Typescript (photocopy). Vita. "Major subject: Electrical Engineering." |
| Physical Description: | vi, 85 leaves : illustrations ; 29 cm |
| Bibliography: | Includes bibliographical references. |