ICASSP 2021 Acoustic Echo Cancellation Challenge

ICASSP 2021声学回声消除挑战旨在刺激声学回声消除 (AEC) 领域的研究，该领域是语音增强的重要组成部分，并且仍然是音频通信和会议系统中的首要问题。最近的许多AEC研究报告了在训练和测试样本来自相同基础分布的合成数据集上的良好性能。但是，AEC的性能通常会在真实记录中显着下降。此外，在现实环境中存在背景噪声和混响的情况下，大多数常规客观指标 (例如回声回波损耗增强 (ERLE) 和语音质量的感知评估 (PESQ)) 与主观语音质量测试没有很好的相关性。在这个挑战中，我们开源了两个大型数据集，以在单谈和双谈场景下训练AEC模型。这些数据集包括来自2,500多个真实音频设备和真实环境中的人类扬声器以及合成数据集的记录。我们开源了两个大型测试集，我们开源了一个在线主观测试框架，供研究人员快速测试他们的结果。这项挑战的获胜者将根据在所有不同的单谈和双谈场景中获得的平均平均意见得分 (MOS) 进行选择。

The ICASSP 2021 Acoustic Echo Cancellation (AEC) Challenge aims to stimulate research in the field of acoustic echo cancellation (AEC), which is a critical component of speech enhancement and remains a core challenge in audio communication and conferencing systems. Many recent AEC studies have reported promising performance on synthetic datasets where training and test samples are drawn from the same underlying distribution. However, the performance of AEC models typically degrades significantly in real-world recorded audio. Furthermore, in realistic environments with background noise and reverberation, most conventional objective metrics—including Echo Return Loss Enhancement (ERLE) and Perceptual Evaluation of Speech Quality (PESQ)—exhibit poor correlation with subjective speech quality assessments. In this challenge, we open-sourced two large-scale datasets for training AEC models under both single-talk and double-talk scenarios. These datasets include recordings from over 2,500 real audio devices, human speakers in real-world environments, as well as synthetic datasets. We additionally open-sourced two large-scale test sets and an online subjective test framework to enable researchers to quickly evaluate their experimental results. The winner of this challenge will be selected based on the average Mean Opinion Score (MOS) across all distinct single-talk and double-talk scenarios.

搜集汇总

背景与挑战

背景概述

ICASSP 2021声学回声消除挑战数据集包含来自2,500多个真实音频设备和环境的记录，以及合成数据，用于训练和测试声学回声消除模型。该数据集旨在解决真实环境中AEC性能下降的问题，并提供了一个在线主观测试框架来评估模型效果。