The Computer Enhancement of Speech Signals
Abstract
Available from UMI in association with The British Library. In this thesis the problem of enhancing speech which has been corrupted by additive noise is considered. Stationary as well as non-stationary, narrowband as well as wideband noise signals are considered and various algorithms developed to reduce their effects. The performance of the developed approaches is analysed in terms of SNR improvement and the effects they have on the intelligibility and quality of the processed speech signal. A single channel approach was developed based on the Wiener theory of optimal filtering. To compute the optimal filter, innovations models for the noise and the "signal plus noise" need to be estimated using the Recursive Extended Least Squares (RELS) algorithm. The optimal filter computation also requires the solution of a single Diophantine equation. Multiple channel approaches are also considered and analysed. A two channel approach is developed and its performance is extensively studied. The effect of the algorithm on a speech recognition system is evaluated. The theory for multiple channel approaches is generalised and the optimum multivariable filter formulation is derived for any system where more than one microphone is used. The tracking performance of the least mean squares (LMS) algorithm is improved by analysing it as a control system hence simplifying the understanding of the engineering concepts of such algorithms. In conclusion to this thesis, it is shown that multiple channel approaches considerably outperform those of single channel approaches. However, there remain difficulties when experiments are performed in realistic environments due to reflections and reverberations.
- Publication:
-
Ph.D. Thesis
- Pub Date:
- 1991
- Bibcode:
- 1991PhDT.......131D
- Keywords:
-
- NOISE;
- OPTIMAL FILTER;
- Engineering: Electronics and Electrical; Physics: Acoustics; Computer Science