CNN-based noise reduction for multi-channel speech enhancement system with discrete wavelet transform (DWT)...

Area of Science:

• Signal Processing
• Artificial Intelligence
• Acoustics

Background:

• Multi-channel speech enhancement (MCSE) systems are crucial for improving speech quality in noisy environments.
• Existing algorithms often struggle with low signal-to-noise ratio (SNR) conditions and high-frequency environmental noises.
• Effective noise reduction is vital for speech devices operating in diverse acoustic settings.

Purpose of the Study:

• To evaluate a novel multi-channel speech enhancement (MCSE) system for noise reduction in stationary and non-stationary environments.
• To compare the performance of an existing MCSE system (BAV-MCSE) with a proposed DWT-CNN-MCSE system across various SNR levels (-10 dB to 20 dB).
• To assess the effectiveness of discrete wavelet transform (DWT) and convolution neural network (CNN) in enhancing noisy speech signals.

Main Methods:

• Experiments utilized the AURORA and LibriSpeech datasets, featuring diverse environmental noises.
• The existing BAV-MCSE system employs beamforming, adaptive noise reduction, and voice activity detection.
• The proposed DWT-CNN-MCSE system integrates discrete wavelet transform (DWT) preprocessing with a convolution neural network (CNN) for denoising.

Main Results:

• The BAV-MCSE achieved a high word recognition rate (WRR) of 93.77% at 20 dB SNR but averaged only 5.64% at -10 dB SNR.
• The proposed DWT-CNN-MCSE system demonstrated superior performance at low SNRs, achieving a WRR of 70.55% at -10 dB SNR.
• The DWT-CNN-MCSE system showed the most significant improvement, with a 64.91% WRR increase at -10 dB SNR.

Conclusions:

• The DWT-CNN-MCSE system offers substantial improvements in speech enhancement, particularly under low SNR conditions.
• This novel approach effectively filters environmental noises, enhancing speech intelligibility in challenging acoustic scenarios.
• The integration of DWT and CNN provides a robust solution for advanced multi-channel speech enhancement.