Deep learning-based stereophonic acoustic echo suppression without decorrelation.

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Stereophonic acoustic echo cancellation (AEC) faces challenges due to correlated signals from stereo loudspeakers, leading to nonuniqueness issues.
• Existing decorrelation methods can degrade audio quality and stereophonic spatial perception.

Purpose of the Study:

• To propose a novel method for stereophonic AEC that avoids decorrelation procedures.
• To maintain audio quality and stereophonic spatial perception while effectively suppressing echo.

Main Methods:

• A convolutional recurrent network (CRN) was developed, integrating convolutional neural networks (CNNs) for feature extraction and recurrent neural networks (RNNs) for temporal modeling.
• The CRN estimates a nonlinear gain applied to the microphone signal's complex spectrum to derive near-end speech.
• Input features utilize the magnitude spectra of far-end signals directly, bypassing decorrelation preprocessing.

Main Results:

• The proposed CRN algorithm demonstrated superior performance compared to traditional methods like normalized least-mean square (NLMS) and Wiener filters.
• Outperformance was particularly notable in challenging acoustic conditions, including low signal-to-echo ratios and high reverberation times (RT60).
• The method successfully maintained both audio quality and stereophonic spatial perception.

Conclusions:

• The CRN-based approach offers an effective solution for stereophonic AEC, overcoming limitations of traditional methods.
• This technique provides a robust alternative for echo suppression in complex acoustic environments, preserving crucial audio characteristics.