DCASE2022 Task 3: Synthetic SELD Audio Mixtures for Baseline Training

1200 one-minute synthetic spatial audio mixtures serve as supplementary training material for the DCASE2022 Challenge. Archontis Politis and colleagues from Tampere University generated the data using real room impulse responses from 9 unique rooms and sound samples from the FSD50K dataset. The dataset supports two 4-channel spatial formats and features a maximum polyphony of 2, with a provided training and testing split.

Use Cases

• Training baseline SELD models based on synthetic spatial audio mixtures.
• Comparing model performance on purely synthetic conditions versus mixed synthetic-real data based on the provided train/test split.
• Fine-tuning models pre-trained on real spatial recordings using the synthetic audio data.
• Researching audio spatialization techniques based on the use of measured room impulse responses from Tampere University.

Strengths

• Contains 1200 one-minute long audio recordings, providing substantial training material.
• Spatialized in 9 unique rooms using measured room impulse responses, offering acoustic variety.
• Supports two 4-channel spatial formats (FOA and MIC), increasing utility for different hardware setups.
• Includes a defined training and testing split, facilitating controlled experimentation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.
• Data may reflect bias inherent to the source FSD50K dataset and the specific rooms at Tampere University.

Provenance

Source

Archontis Politis, Tampere University, generated for the DCASE2022 Challenge.

Collection Method

Synthetic generation using sound samples from FSD50K spatialized with measured room impulse responses from Tampere University.

Geography

Room impulse responses were measured at Tampere University (TAU).