Low-dimensional recurrent neural network-based Kalman filter for speech enhancement.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Speech enhancement is crucial for clear audio communication.
• Traditional methods struggle with non-Gaussian noise and computational efficiency.
• Recurrent Neural Networks (RNNs) offer potential for adaptive signal processing.

Purpose of the Study:

• To develop a novel RNN-based Kalman filter for robust speech enhancement.
• To improve estimation accuracy and computational speed compared to existing methods.
• To address challenges posed by non-Gaussian noise in speech signals.

Main Methods:

• Utilized a noise-constrained least squares estimate for parameter estimation.
• Employed a recurrent neural network (RNN) to model autoregressive speech signal parameters.
• Integrated the RNN with a Kalman filter for speech signal recovery.
• Ensured global asymptotic stability of the RNN for noise-constrained estimates.

Main Results:

• The proposed algorithm demonstrated robust performance in non-Gaussian noise.
• Achieved minimized estimation error for Kalman filter parameters under non-Gaussian conditions.
• Exhibited significantly faster computation speeds due to a low-dimensional model feature.
• Simulation results confirmed effective noise reduction and good overall performance.

Conclusions:

• The RNN-based Kalman filter offers an efficient and effective solution for speech enhancement.
• The noise-constrained approach enhances robustness against non-Gaussian noise.
• This method presents a promising advancement in real-time speech processing applications.