Convolutional fusion network for monaural speech enhancement.

Area of Science:

• Signal Processing
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional neural network (CNN) based methods achieve state-of-the-art results in monaural speech enhancement.
• Conventional encoder-decoder networks use large kernels, leading to low parameter efficiency.
• Existing methods like ShuffleNet address this but may lose inter-channel dependency information.

Purpose of the Study:

• To propose a new Convolutional Fusion Network (CFN) for monaural speech enhancement.
• To improve model performance, inter-channel dependency, information reuse, and parameter efficiency.
• To address limitations of conventional encoder-decoder networks and ShuffleNet.

Main Methods:

• Introduced a novel Group Convolutional Fusion Unit (GCFU) using standard and depth-wise separable CNNs.
• Implemented a parallel network structure where full input sequences are processed by two CNNs, with outputs shuffled and concatenated.
• Utilized an intra skip connection mechanism within the encoder and decoder to enhance information flow.

Main Results:

• The proposed CFN demonstrates improved performance in monaural speech enhancement.
• The method effectively exploits inter-channel dependency, enhancing signal reconstruction.
• Experimental results show superior performance compared to three recent baseline methods.

Conclusions:

• The novel CFN effectively enhances monaural speech.
• The proposed GCFU and network architecture improve parameter efficiency and inter-channel dependency.
• This work offers a promising advancement in deep learning-based speech enhancement techniques.