REGULARIZED OPTIMIZATION WITH SPECTRAL SMOOTHING FOR SPEECH SPECTRAL ESTIMATION
Karsten Sørensen, S.V. Andersen · 2003
This paper presents a minimization approach based on regularized optimization [1] for use in spectral envelope estimation for speech in the presence of noise. The objective function that is minimized consists of the sum of a data fitting term and a linear combination of the following regularization terms: Minimum perturbation energy, minimum estimate energy, peak preservation, and spectral smoothing. Regularization weights are used to control the tradeoff between the regularization terms such that the estimates obtain a lower bias and low variance, while preserving a shape that is natural for speech.