Characterization of Deep Learning-Based Speech-Enhancement Techniques in Online Audio Processing Applications.

Area of Science:

• Audio Signal Processing
• Machine Learning
• Digital Communications

Background:

• Deep learning models show promise for speech enhancement in offline scenarios.
• Evaluating these models for online, real-time audio processing is crucial for practical applications.

Purpose of the Study:

• To assess the online applicability of state-of-the-art deep learning speech enhancement techniques.
• To characterize model performance concerning input segment length, noise, interference, and reverberation.

Main Methods:

• Systematic evaluation of three popular models (MetricGAN+, SFM-ML, Demucs-Denoiser) using the Speechbrain framework.
• Analysis of output signal-to-interference ratio, response time, and memory usage.
• Investigation of the impact of varying input lengths and audio degradation levels.

Main Results:

• Performance metrics (e.g., signal-to-interference ratio, response time, memory usage) are influenced by input segment length.
• Different models exhibit varying trade-offs between enhancement quality and online processing efficiency.
• Audio conditions (noise, interference, reverberation) significantly affect model performance in online settings.

Conclusions:

• The study provides the first comprehensive characterization of deep learning speech enhancement models for online use.
• Recommendations for future research are proposed to optimize models for real-time digital voice communication.
• Understanding these trade-offs is essential for selecting and deploying effective speech enhancement solutions.