Low-latency monaural speech enhancement with deep filter-bank equalizer.

Area of Science:

• Signal Processing
• Artificial Intelligence
• Acoustics

Background:

• Low-latency speech enhancement is crucial for real-time applications like hearing aids and mobile phones.
• Traditional algorithms often face a trade-off between performance and latency.
• Improving speech quality while maintaining minimal delay remains a significant challenge.

Purpose of the Study:

• To introduce a novel deep filter-bank equalizer (DeepFBE) framework for low-latency speech enhancement.
• To integrate deep learning for subband noise reduction and filter mapping to achieve high performance with minimal latency.
• To validate the effectiveness of DeepFBE against traditional low-latency speech enhancement methods.

Main Methods:

• A deep learning-based subband noise reduction network was trained with a small frame shift (≤4ms) to generate adaptive digital filter gains.
• A second deep learning network implicitly shortened the adaptive digital filter to further reduce latency.
• The proposed framework reconstructed enhanced speech without the conventional overlap-add method.

Main Results:

• The DeepFBE framework achieved a latency of only 4ms.
• Experimental results on the WSJ0-SI84 corpus showed significantly better performance compared to traditional low-latency algorithms across objective metrics.
• Listening tests confirmed superior speech quality achieved by the DeepFBE approach.

Conclusions:

• The proposed DeepFBE framework effectively achieves high-performance, low-latency speech enhancement.
• Deep learning integration in filter-bank equalization offers a promising direction for real-time speech processing applications.
• DeepFBE provides a substantial improvement in speech quality and performance for devices requiring minimal delay.