Frequency-domain beamformers using conjugate gradient techniques for speech enhancement.

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Frequency-domain minimum-variance-distortionless-response (MVDR) beamformers are crucial for signal enhancement.
• Conventional methods often require explicit or implicit autocorrelation matrix inversion, which can be computationally intensive and sensitive to data limitations.

Purpose of the Study:

• To propose novel algorithms for implementing MVDR beamformers that avoid matrix inversion.
• To apply these algorithms to the problem of speech enhancement.
• To analyze the convergence properties and performance of the proposed methods.

Main Methods:

• Development of multiple-iteration constrained conjugate gradient (MICCG) and single-iteration constrained conjugate gradient (SICCG) algorithms.
• Derivation using Lagrange multipliers and conjugate gradient techniques.
• Implementation avoiding explicit or implicit autocorrelation matrix inversion.

Main Results:

• Theoretical convergence guarantees for both MICCG and SICCG algorithms.
• MICCG provides a finite sequence of estimates converging to the MVDR solution, outperforming conventional estimators for limited data.
• SICCG offers sample-by-sample updating with asymptotic convergence to the MVDR solution.
• Demonstrated performance on synthetic and real acoustic vector sensor array data.

Conclusions:

• The proposed MICCG and SICCG algorithms offer efficient and robust alternatives for MVDR beamforming in speech enhancement.
• These methods effectively address the limitations of traditional approaches by avoiding matrix inversion.
• The algorithms show promise for applications requiring adaptive and high-performance beamforming.