ChannelSet: a composite dataset of diverse acoustic environments

We introduce ChannelSet, a dataset which provides a launchpad for exploring the extraneous acoustic information typically suppressed or ignored in audio tasks such as automatic speech recognition. We combined components of existing publicly available datasets to encompass broad variability in recording equipment, microphone position, room or surrounding acoustics, event density (i.e., how many audio events are present), and proportion of foreground and background sounds. Source datasets include: the CHiME-3 background dataset, CHiME-5 evaluation dataset, AMI meeting corpus, Freefield1010, and Vystadial2016. ChannelSet includes 13 classes spanning various acoustic environments: Indoor_Commercial_Bus, Indoor_Commercial_Cafe, Indoor_Domestic, Indoor_Meeting_Room1, Indoor_Meeting_Room2, Indoor_Meeting_Room3, Outdoor_City_Pedestrian, Outdoor_City_Traffic, Outdoor_Nature_Birds, Outdoor_Nature_Water, Outdoor_Nature_Weather, Telephony_CZ, and Telephony_EN. Each sample is between 1 and 10 seconds in duration. Each class contains 100 minutes of audio, for a total of 21.6 hours, split into separate test (20%) and train (80%) partitions. Download includes scripts, metadata, and instructions for producing ChannelSet from source datasets