Layer modified residual Unet++ for speech enhancement using Aquila Black widow optimizer algorithm.

Area of Science:

• Signal Processing
• Artificial Intelligence
• Machine Learning

Background:

• Speech enhancement is crucial but computationally demanding.
• Deep learning models struggle with diverse environmental noises, requiring robust systems.
• Existing methods face challenges in efficiency and quality.

Purpose of the Study:

• To develop a lightweight and efficient deep learning model for environmental speech enhancement.
• To introduce a novel heuristic-inspired model for robust noise removal.
• To improve the quality of speech signals degraded by various noises.

Main Methods:

• Utilized Short-Time Fourier Transform (STFT) to convert noisy speech to spectrograms.
• Developed Layer Modified Residual Unet++ (LMResUnet++) with atrous convolution for multi-scale feature extraction.
• Employed Aquila Black Widow Optimization (ABWO) for hyperparameter tuning and model optimization.
• Restored enhanced speech via Inverse STFT.

Main Results:

• The LMResUnet++ model demonstrated superior performance in speech enhancement.
• Achieved significant improvements in Perceptual Evaluation of Speech Quality (PESQ) scores.
• Outperformed existing models like DeepUnet, MTCNN, STCNN, and ResUnet++ by notable margins (7.93% to 1.90%).

Conclusions:

• The proposed LMResUnet++ offers an efficient and robust solution for environmental speech enhancement.
• The hybrid optimization approach enhances model compactness and performance.
• This deep learning design effectively removes noise while preserving speech quality.