ICASSP 2021 Acoustic Echo Cancellation Challenge: Datasets, Testing Framework, and Results

Abstract

The ICASSP 2021 Acoustic Echo Cancellation Challenge is intended to stimulate research in the area of acoustic echo cancellation (AEC), which is an important part of speech enhancement and still a top issue in audio communication and conferencing systems. Many recent AEC studies report good performance on synthetic datasets where the train and test samples come from the same underlying distribution. However, the AEC performance often degrades significantly on real recordings. Also, most of the conventional objective metrics such as echo return loss enhancement (ERLE) and perceptual evaluation of speech quality (PESQ) do not correlate well with subjective speech quality tests in the presence of background noise and reverberation found in realistic environments. In this challenge, we open source two large datasets to train AEC models under both single talk and double talk scenarios. These datasets consist of recordings from more than 2,500 real audio devices and human speakers in real environments, as well as a synthetic dataset. We open source two large test sets, and we open source an online subjective test framework for researchers to quickly test their results. The winners of this challenge will be selected based on the average Mean Opinion Score (MOS) achieved across all different single talk and double talk scenarios.

Personalized Acoustic Echo Cancellation for Full-duplex Communications

“…The best results in each column and the best overall results are bolded and underlined, respectively. Samples of the processed audio clips can be found on our demo page 1 .…”

“…Deep neural network based AEC algorithms flourish with the recent series of AEC-Challenges [1,2,3]. Westhausen et al [4] adopt a dual-signal transformation LSTM network (DTLN) and use both the microphone signal and the far-end signal to predict the near-end speech in an end-to-end fashion.…”

Y$^2$-Net FCRN for Acoustic Echo and Noise Suppression

“…The "other degradation DMOS" test is dubbed as "NE", describing quality of NE speech. More details can be found in [24].…”

“…The model is trained with synthetic files from the database provided by Microsoft for the Interspeech 2021 AEC Challenge [24], in the following labeled as Dsyn. The audio files were created with varying conditions, including single-talk, double-talk, both NE and FE noise, as well as simulated nonlinear loudspeaker distortions.…”

“…The IRs for the FE echoes are generated using Microsoft's Project Triton [26] technology with reverberation times T60 between 0.2 s and 1.2 s. For 80% of the files the FE signal is also subject to one of various nonlinear functions before applying the IR to simulate loudspeaker distortions. A more in-depth description can be found in [24]. A number of 500 files of Dsyn were created with unseen speakers and IRs, and are used as preliminary synthetic testset D test,pre syn for instrumental measurements.…”