A convolutional recurrent neural network with attention framework for speech separation in monaural recordings.

Area of Science:

• Signal Processing
• Artificial Intelligence
• Machine Learning

Background:

• Monaural speech separation often yields unsatisfactory results due to limitations of single network architectures.
• Existing methods struggle to effectively learn features for high-quality speech separation.

Purpose of the Study:

• To propose a novel Convolutional Recurrent Neural Network with Attention (CRNN-A) framework for enhanced monaural speech separation.
• To fuse the complementary strengths of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) for improved feature learning.

Main Methods:

• A CNN front-end with dual convolution kernels extracts time- and frequency-domain features from spectrograms.
• Features are concatenated and processed through convolutional layers before being fed to an RNN back-end.
• An attention mechanism is integrated to focus on relationships between feature maps, optimizing separation performance.

Main Results:

• The CRNN-A framework demonstrated superior performance compared to baseline RNN and other popular speech separation methods.
• Evaluations on the MIR-1K dataset showed significant improvements in Global Normalised Source-to-Distortion Ratio (GNSDR), Global Source-to-Interference Ratio (GSIR), and Global Source-to-Artifacts Ratio (GSAR).

Conclusions:

• The proposed CRNN-A framework effectively integrates CNN and RNN capabilities, enhanced by an attention mechanism, for superior speech separation.
• This framework offers a promising approach for speech separation, speech enhancement, and related audio processing fields.