A method for enhancing speech and warning signals based on parallel convolutional neural networks in a noisy...

Area of Science:

• Signal Processing
• Artificial Intelligence
• Auditory Neuroscience

Background:

• Digital hearing aids amplify sound but may remove crucial warning sounds, posing risks to users.
• Existing noise reduction techniques can inadvertently suppress important auditory alerts.

Purpose of the Study:

• To develop a deep learning model for distinguishing warning sounds in speech signals contaminated with noise.
• To enhance the intelligibility of speech and warning sounds by effectively removing noise.

Main Methods:

• An adaptive convolution filter was developed and applied to a PCNNs (Pulse Coupled Neural Networks) model for time and frequency domain analysis.
• The PCNNs model classifies the presence or absence of warning sounds.
• A CEDN (Context-aware End-to-End Denoising Network) model was employed to improve sound intelligibility based on warning sound classification.

Main Results:

• The proposed PCNNs model with multiplicative filters demonstrated efficiency in analyzing complex sound signals.
• The CEDN model successfully enhanced the intelligibility of speech and warning sounds.
• The combined model showed high training rates, low error rates, and stable performance.

Conclusions:

• The PCNN model integrated with the proposed filter achieved superior training and error rates, indicating robust performance.
• While the CEDN model improved recognition of speech and warning sounds, its effectiveness decreased with higher noise ratios.
• Further research is needed to optimize performance in extremely noisy conditions.