STARSS23: Sony-TAu Realistic Spatial Soundscapes 2023

STARSS23 contains multichannel recordings of sound scenes in various rooms and environments, with temporal and spatial annotations of prominent sound events. The dataset was collected by Tampere University in Finland and SONY in Japan between September 2021 and February 2023, serving as the development dataset for the DCASE 2023 Challenge. It includes over 4 hours of additional recordings compared to its predecessor, featuring 360° videos and distance labels for sound events.

Use Cases

• Training sound event localization models based on multichannel audio recordings and spatial annotations.
• Developing multimodal audio-visual models based on the 360° videos aligned with the audio recordings.
• Benchmarking detection algorithms for real-world sound scenes based on the dataset's realistic, non-synthetic nature.
• Researching distance estimation for sound events based on the newly added distance labels in centimeters.

Strengths

• Contains real-world recordings, avoiding pitfalls of synthetic scene generation mentioned in the description.
• Includes over 4 hours of additional audio data compared to the previous STARSS22 version.
• Provides multimodal data with 360° videos spatially and temporally aligned to most audio clips.
• Annotations combine human annotators for event activity and optical tracking for spatial positions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and total dataset size are unknown, which may limit suitability assessment.
• Data may reflect geographic bias inherent to collection in only two countries (Finland and Japan).

Provenance

Source

Tampere University (Audio Research Group) and SONY.

Collection Method

Multichannel recordings collected using a similar setup and annotation procedure in real rooms and environments.

Time Range

First round: September 2021 to January 2022. Second round: November 2022 to February 2023.

Geography

Tampere, Finland and Tokyo, Japan.