SpatialSounds Dataset Overview

Dataset Contents

AudioSet (non-reverberant audio sources)

• Dataset Type: Mono/Binaural/Ambisonics
• Dataset Structure:
  • Balanced train and Evaluation sets
  • Unbalanced train set requires reference to Official AudioSet
• Metadata: Available for download from metadata
• Weight Generation: See weights-generation or use the provided weight files

Reverberation

• Dataset Structure:
  • train_reverberation.json and eval_reverberation.json
  • binaural and mono folders
• Download Link: mp3d_reverberation

SpatialSoundQA Dataset

• Dataset Content: Training data for different stages of training BAT models
• Download Link: SpatialSoundQA

Data Generation Method

• Spatial Audio Generation: Use scipy.signal.fftconvolve or torchaudio.functional.fftconvolve to generate spatial audio from mono recordings

TODO

1. Upload QA evaluation set

License

• License Type: CC BY-NC 4.0