TUT Sound Events 2018: Ambisonic Recordings with Spatial Metadata

Recordings from 2018 consist of real-life first order Ambisonic (FOA) format audio with stationary point sources. The dataset was generated by collecting impulse responses in a university corridor using an Eigenmike spherical microphone array and a moving loudspeaker. It contains three sub-datasets for overlapping sound events, with metadata including temporal onset/offset and spatial coordinates for events sourced from the UrbanSound8K dataset.

Use Cases

• Training models for sound event detection and localization based on annotated temporal and spatial metadata.
• Developing and evaluating algorithms for Ambisonic audio processing and dereverberation.
• Researching overlapping sound source separation in real-world reverberant environments.
• Benchmarking spatial audio machine learning models using provided cross-validation splits.

Strengths

• Provides real-world impulse responses collected with a high-fidelity Eigenmike array at multiple elevations and distances.
• Includes detailed metadata for each sound event: temporal onset/offset, azimuth, elevation, and distance.
• Structured for machine learning with three distinct sub-datasets for overlapping events and three cross-validation splits per sub-dataset.

Limitations

• The exact total number of recordings, rows, and file size is not specified across the provided sources.
• Metadata completeness is unclear as column names are listed as 'unknown'.
• The dataset uses a subset of sound event classes from UrbanSound8K, excluding air_conditioner and children_playing.

Provenance

Source

Tampere University of Technology (TUT), created by Sharath Adavanne.

Collection Method

Impulse responses were collected by moving a Genelec G Two loudspeaker playing a maximum length sequence around an Eigenmike array in a corridor.

Time Range

2018

Geography

Recorded in a corridor inside Tampere University of Technology.