TUT Sound Events 2018: Ambisonic Recordings with Spatial Metadata

A 2018 dataset from Tampere University of Technology contains simulated anechoic Ambisonic recordings of 11 sound event classes, such as coughing, door knocks, and speech. It includes three sub-datasets with up to three temporally overlapping sound events, each with cross-validation splits totaling 240 training and 60 test recordings of about 30 seconds each. Each recording is paired with metadata specifying the event's name, timing, and spatial location in azimuth, elevation, and distance.

Use Cases

• Training sound event localization models based on spatial coordinate metadata (azimuth, elevation, distance).
• Developing detection algorithms for overlapping sound events based on the sub-datasets with up to three simultaneous events.
• Benchmarking spatial audio processing systems using the provided cross-validation splits and standardized 30-second recordings.

Strengths

• Provides spatial metadata (azimuth, elevation, distance) for each sound event, enabling precise localization tasks.
• Includes three structured sub-datasets with 1, 2, or 3 overlapping events, facilitating research on source separation.
• Offers cross-validation splits with 240 training and 60 test recordings per split, supporting reproducible model evaluation.

Limitations

• Data is simulated and anechoic, which may not fully represent real-world acoustic environments with reverberation.
• Row count and total file size are unknown, which may limit suitability assessment for large-scale training.
• Column-level documentation is absent; field semantics must be inferred from the description after download.

Provenance

Source

Tampere University of Technology, created by Sharath Adavanne.

Collection Method

Simulated recordings using isolated sound events from the DCASE 2016 task 2 dataset, placed on a spatial grid.

Time Range

2018