AEC in A Netshell: on Target and Topology Choices for FCRN Acoustic Echo Cancellation

Abstract

Acoustic echo cancellation (AEC) algorithms have a long-term steady role in signal processing, with approaches improving the performance of applications such as automotive hands-free systems, smart home and loudspeaker devices, or web conference systems. Just recently, very first deep neural network (DNN)-based approaches were proposed with a DNN for joint AEC and residual echo suppression (RES)/noise reduction, showing significant improvements in terms of echo suppression performance. Noise reduction algorithms, on the other hand, have enjoyed already a lot of attention with regard to DNN approaches, with the fully convolutional recurrent network (FCRN) architecture being among state of the art topologies. The recently published impressive echo cancellation performance of joint AEC/RES DNNs, however, so far came along with an undeniable impairment of speech quality. In this work we will heal this issue and significantly improve the near-end speech component quality over existing approaches. Also, we propose for the first time-to the best of our knowledge-a pure DNN AEC in the form of an echo estimator, that is based on a competitive FCRN structure and delivers a quality useful for practical applications.

Deep Residual Echo Suppression and Noise Reduction: A Multi-Input FCRN Approach in a Hybrid Speech Enhancement System

“…We use three measures to evaluate all approaches: Speech quality in terms of wideband PESQ MOS LQO [26,27], SNR improvement (∆SNR) in [dB] for noise reduction, and echo suppression by echo return loss enhancement (ERLE) in [dB] computed as in [8].…”

“…The following rows provide two approaches as baseline comparisons: First, a single-stage 'all-in-one' approach trains the FCRN to perform all tasks (AEC+RES+NR) at once, as proposed in [6] using direct estimation and further elaborated in [8]. As for the Kalman filter, the only available inputs to this approach are the microphone signal Y and reference signal X (for brevity denoted as vectors without frame and bin index here).…”

“…The acoustic setup is outlined in Figure 1 and shall cover typical single-and double-talk scenarios. Following the procedure described in [8,6], the TIMIT dataset [15] is used to set up far-end speech x(n) and near-end speech s(n) at a sampling rate of 16 kHz. At the microphone, near-end speech s(n) is superimposed with background noise n(n) and echo signals d(n).…”

“…Echoes d(n) are generated by imposing loudspeaker nonlinearities [6] on far-end signals x(n) and applying impulse responses (IRs) created with the image method [16]. As in related work [8], IRs are set to a length of 512 samples with reverberation times T60 ∈ {0.2, 0.3, 0.4} s for training and validation, and 0.2 s for test mixtures. Background noise n(n) from the QUT dataset [17] is used for training and validation.…”

“…In this typical two-stage arrangement for speech enhancement, the application of a DNN as second stage (RES and NR) gained increasing attention with early investigations of feed-forward networks [1,2], convolutional networks bringing further improvements more recently [3,4], some even being fully synergistic with the first stage [5], and many more. In the meantime, also fully learned deep AEC approaches were proposed, where a single network incorporates the tasks of AEC, RES, and NR, e.g., [6,7] or further investigated in [8]. Although showing an impressive suppression performance, however, these are often accompanied by some near-end speech degradation and-at least for now-hybrid approaches are still one step ahead as can be seen with the leading model of the AEC Challenge on ICASSP 2021 being a hybrid approach [9].…”