TUT Sound Events 2018: Simulated Ambisonic Recordings with Overlapping Sources

240 training and 60 test recordings, each about 30 seconds long, simulate first-order Ambisonic audio with up to three overlapping sound events. The dataset, created by Sharath Adavanne at Tampere University, uses 11 sound event classes from the DCASE 2016 task 2 dataset, spatially placed in a simulated 10x8x4 meter room. Each recording includes metadata with event timings and spatial coordinates in azimuth, elevation, and distance.

Use Cases

• Training sound event localization models based on simulated first-order Ambisonic recordings with spatial metadata.
• Developing algorithms for detecting overlapping sound events based on datasets with up to three temporally overlapping sources.
• Benchmarking acoustic scene analysis systems using synthetic impulse responses and reverberant conditions described in the dataset.
• Researching spatial audio perception and source separation using data with annotated azimuth, elevation, and distance coordinates.

Strengths

• Provides three distinct sub-datasets for training models with one, two, or three temporally overlapping sound events.
• Includes 240 training and 60 test recordings per cross-validation split, each about 30 seconds long.
• Metadata for each recording specifies sound event class, onset/offset times, and precise spatial coordinates (azimuth, elevation, distance).
• Sound events are placed using a defined spatial grid with 10-degree azimuth resolution and elevation between -60 and 60 degrees.

Limitations

• Row count and total file size are unknown, which may limit suitability assessment for large-scale projects.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data is entirely synthetic, generated using the image source method, which may not fully capture real-world acoustic variability.

Provenance

Source

Tampere University

Collection Method

Synthetic generation using the image source method, with isolated sound events sourced from the DCASE 2016 task 2 dataset.