A Laplacian-based MMSE estimator for speech enhancement

Area of Science:

• Digital Signal Processing
• Acoustic Signal Processing

Background:

• Speech enhancement aims to improve speech quality by reducing noise.
• Traditional methods often assume Gaussian distributions for speech and noise, which may not be optimal.
• Accurate modeling of the underlying signal distributions is crucial for effective enhancement.

Purpose of the Study:

• To develop and evaluate optimal estimators for the magnitude spectrum in speech enhancement.
• To investigate the impact of different statistical models (Laplacian vs. Gaussian) on speech enhancement performance.
• To derive a Minimum Mean Square Error (MMSE) estimator under speech presence uncertainty using a Laplacian model.

Main Methods:

• Analytical derivation of magnitude spectrum estimators in the MMSE sense.
• Modeling clean speech DFT coefficients using a Laplacian distribution.
• Modeling noise DFT coefficients using a Gaussian distribution.
• Developing an MMSE estimator incorporating speech presence uncertainty and a Laplacian model.

Main Results:

• The Laplacian-based MMSE estimator demonstrated reduced residual noise compared to the traditional Gaussian-based MMSE estimator.
• The study confirmed that the assumed distribution of Discrete Fourier Transform (DFT) coefficients significantly impacts enhanced speech quality.
• The derived MMSE estimator under speech presence uncertainty with a Laplacian model showed improved performance.

Conclusions:

• The choice of statistical distribution for DFT coefficients is critical for effective speech enhancement.
• Laplacian-based models offer superior performance in reducing residual noise for speech enhancement compared to Gaussian models.
• The developed MMSE estimator provides a more robust approach to speech enhancement, especially under conditions of speech presence uncertainty.