Two-stage Deep Learning for Noisy-reverberant Speech Enhancement

Area of Science:

• Signal Processing
• Artificial Intelligence
• Acoustics

Background:

• Real-world speech is often distorted by background noise and room reverberation.
• These distortions degrade speech intelligibility and quality, impacting speech recognition systems.

Purpose of the Study:

• To propose a novel two-stage strategy for enhancing speech corrupted by both noise and reverberation.
• To improve speech intelligibility and quality in challenging acoustic environments.

Main Methods:

• A two-stage deep neural network (DNN) approach for sequential denoising and dereverberation.
• Development of a new objective function incorporating clean phase for improved spectral and phase estimation.
• Joint training of the two-stage model using the proposed objective function.

Main Results:

• The proposed algorithm substantially improves objective metrics for speech intelligibility and quality.
• The two-stage enhancement system significantly outperforms previous one-stage methods.
• Enhanced speech quality and intelligibility demonstrated through systematic evaluations.

Conclusions:

• The proposed two-stage strategy effectively addresses combined noise and reverberation in speech.
• The novel objective function aids in more accurate spectral and phase reconstruction.
• This method offers a significant advancement for speech enhancement applications.