RESIDUAL RECURRENT NEURAL NETWORK FOR SPEECH ENHANCEMENT.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Current speech enhancement models often rely on spectrograms, leading to costly transformations and loss of phase information.
• Prior convolutional network approaches struggle with memory-intensive operations and aliasing from upsampling.

Purpose of the Study:

• To develop an end-to-end fully recurrent neural network for single-channel speech enhancement.
• To overcome limitations of existing models by efficiently capturing long-range temporal dependencies without information loss.

Main Methods:

• An hourglass-shaped recurrent neural network architecture was employed.
• The network reduces feature resolution to capture temporal dependencies efficiently.
• Residual connections were utilized to prevent gradient decay and enhance model generalization.

Main Results:

• The proposed recurrent neural network model demonstrated superior performance.
• Outperformed state-of-the-art approaches across six quantitative evaluation metrics.
• Successfully addressed issues of phase information loss and computational complexity.

Conclusions:

• The end-to-end recurrent neural network offers an efficient and effective solution for single-channel speech enhancement.
• The hourglass architecture and residual connections contribute to improved performance and generalization.
• This approach represents a significant advancement over traditional spectrogram-based and convolutional methods.