Deep Learning Based Real-time Speech Enhancement for Dual-microphone Mobile Phones

Area of Science:

• Signal Processing
• Artificial Intelligence
• Mobile Communications

Background:

• Mobile speech communication is often degraded by background noise.
• Speech enhancement systems using microphones are common in mobile phones.
• Existing methods may not be optimal for dual-microphone setups.

Purpose of the Study:

• To propose a novel deep learning approach for real-time speech enhancement.
• To develop an efficient system for dual-microphone mobile phones.
• To improve the quality of speech signals in noisy mobile environments.

Main Methods:

• A densely-connected convolutional recurrent network for dual-channel complex spectral mapping.
• Structured pruning technique for model compression.
• Real-time processing for low-latency and memory efficiency.

Main Results:

• The proposed deep learning approach significantly enhances speech quality.
• The system demonstrates superior performance compared to previous dual-channel methods.
• Outperforms a deep learning-based beamformer in speech enhancement.

Conclusions:

• The novel deep learning approach offers effective real-time speech enhancement for dual-microphone mobile phones.
• Model compression techniques enable efficient and low-latency processing.
• This method represents a significant advancement in mobile communication audio quality.