TAU Moving Sound Events 2019: Simulated Ambisonic Audio with Spatial Trajectories

Tampere University's TAU Moving Sound Events 2019 dataset provides simulated anechoic, first-order Ambisonic (FOA) recordings of 11 everyday sound classes moving in 2D spherical space. It consists of three sub-datasets with up to one, two, or three temporally overlapping events, each with structured training and test splits. Each recording includes metadata detailing the event's name, timing, spatial trajectory, angular velocity, and distance from the microphone.

Use Cases

• Training machine learning models for moving sound event detection based on annotated temporal and spatial metadata.
• Developing and benchmarking spatial audio source separation algorithms using recordings with up to three overlapping events.
• Evaluating direction-of-arrival (DOA) estimation methods for sources moving with constant angular velocity on an arc.
• Studying the impact of source distance (1-10 meters) on spatial audio encoding in an anechoic, simulated environment.

Strengths

• Provides structured cross-validation splits, with 240 training and 60 test recordings per sub-dataset, each about 30 seconds long.
• Includes detailed per-recording metadata for sound event name, timing, spatial trajectory (azimuth, elevation), angular velocity, and distance.
• Systematically varies acoustic complexity across three sub-datasets with a maximum of one, two, or three temporally overlapping sound events.

Limitations

• The dataset is entirely synthetic and anechoic, which may not represent real-world acoustic conditions with reverberation and noise.
• Source movement is constrained to arcs at a constant distance from the microphone, limiting trajectory diversity.
• Specific file formats, total size, and exact row counts are not provided in the available metadata.

Provenance

Source

Tampere University, created by Sharath Adavanne.

Collection Method

Synthesized using simulated anechoic first-order Ambisonic (FOA) recordings. Isolated sound events were taken from the DCASE 2016 Task 2 dataset.

Time Range

2019