ICASSP 2021 Acoustic Echo Cancellation Challenge: Integrated Adaptive Echo Cancellation with Time Alignment and Deep Learning-Based Residual Echo Plus Noise Suppression

Abstract

This paper describes a three-stage acoustic echo cancellation (AEC) and suppression framework for the ICASSP 2021 AEC Challenge. In the first stage, a partitioned block frequency domain adaptive filtering is implemented to cancel the linear echo components without introducing the near-end speech distortion, where we compensate the time delay between the far-end reference signal and the microphone signal beforehand. In the second stage, a deep complex U-Net integrated with gated recurrent unit is proposed to further suppress the residual echo components. In the last stage, an extremely tiny deep complex U-Net is trained to suppress non-speech residual components that have not been suppressed completely in the second stage, which can also further increase the echo return loss enhancement (ERLE) without increasing the computational complexity dramatically. Experimental results show that the proposed three-stage framework can get the ERLE higher than 50 dB in both single-talk and double-talk scenarios, and perceptual evaluation of speech quality can be improved about 0.75 in double-talk scenarios. The proposed framework outperforms the AEC-Challenge baseline ResRNN by 0.12 points in terms of the MOS.

Deep Learning-Based Acoustic Echo Cancellation for Surround Sound Systems

“…The least mean square (LMS) algorithm is one of the most widely used adaptation methods in the echo path identification. Among LMS algorithms, the partitioned block frequency domain LMS (PBFDLMS) algorithm is popular for its lower computational complexity than the time-domain LMS algorithms and its lower latency than frequencydomain LMS algorithms, especially when the acoustic echo path is relatively long [14]. To validate the proposed algorithm, the PBFDLMS algorithm with Wiener post-filtering was chosen as the baseline.…”

“…Cheng et al [13] proposed a convolutional recurrent network (CRN) model to estimate the non-linear gain from the magnitude spectra of both the microphone and far-end signals for the stereo AEC, which was then multiplied by the spectrum of the microphone signal to estimate near-end speech. Besides, Peng et al [14] described a threestage AEC and suppression framework for the ICASSP 2021 AEC Challenge, where the partitioned block frequency domain least mean square (PBFDLMS) with a time alignment was firstly implemented to cancel the linear echo components, and two deep learning networks were then proposed to suppress the residual echo and the non-speech residual noise simultaneously. In addition, Zhang et al [15] proposed a neural cascade architecture, including a CRN module and an LSTM module, which is used for joint acoustic echo and noise suppression to address both single-channel and multi-channel AEC problems.…”

A Multi-Stage Acoustic Echo Cancellation Model Based on Adaptive Filters and Deep Neural Networks

“…To demonstrate the superior performance of our multi-stage acoustic echo cancellation model, we conduct a comprehensive comparison with DCGRU22 [43], DTLN [18], and MTFAA [15]. Notably, DCGRU22 and DTLN are ranked as the 5th and 7th models in the ICASSP 2021 Acoustic Echo Cancellation challenge, respectively.…”

A deep complex multi-frame filtering network for stereophonic acoustic echo cancellation

“…Deep learning-based MAEC methods are robust to double-talk situation and can better suppress acoustic echo in non-stationary noise environments. The performance of deep learing-based MAEC methods in real-world environment is verified in AEC chanllenge [17,18], which open large real recording datasets for researchers to test their proposed algorithm in practical situation.…”