microsoft/AEC-Challenge

What it solves

Acoustic echo cancellation (AEC) is a critical but still challenging part of speech enhancement in audio communication. This project supplies researchers with the data and evaluation resources needed to develop and test new AEC algorithms, aiming to push the field forward.

How it works

The project open-sources two large training datasets: real recordings captured from more than 10,000 actual audio devices and human speakers in real environments, plus a synthetic dataset. Both cover single-talk and double-talk scenarios. To measure performance, it provides an online subjective test framework and an objective metric service (including a full-band version of AECMOS) so results can be quickly checked. The challenge also adds a second track for personalized AEC and reduces the algorithmic+buffering latency to 20 ms.

Who it's for

Researchers and engineers working on acoustic echo cancellation, speech enhancement, or audio communication. It is also directly intended for participants in the ICASSP 2023 Acoustic Echo Cancellation Challenge who want to benchmark their algorithms against a common dataset and evaluation pipeline.

Highlights

• Two large open-source datasets: real recordings from over 10,000 devices and a synthetic set, covering both single-talk and double-talk.
• Online subjective test framework and an objective metric service (AECMOS) for quick evaluation.
• Second track for personalized AEC and a reduced latency target of 20 ms.
• Winners selected based on average Mean Opinion Score (MOS) and word accuracy rate.