Speech separation and recognition in real, everyday acoustic environments.
The CHiME challenge series releases conversational, multi-channel and wearable speech recordings made in the places people actually talk — homes, cars, cafés, streets, and more recently through smart glasses — and runs open evaluations on them. The emphasis throughout has been on real recordings rather than simulation, and on tasks that stay close to how the audio would really be captured and used.
This organization hosts the openly licensed CHiME datasets.