spatial-audio-learning/gram-scenes-as2m-16k

GRAM-style binaural scenes from AudioSet-2M, 16 kHz

AudioSet-2M rendered to binaural stereo following the scene-synthesis recipe of GRAM (arXiv:2506.00934), used to pretrain stereo BEATs encoders.

64 shard tars, 829.5 GB, 2,020,706 clips — one scene per clip of AudioSet's unbalanced_train set.

Rendering

Per clip, as in GRAM-T (dataset_functions.py): AudioSet mono -> RMS -14 dBFS -> 10 s pad/truncate -> convolve with a binaural RIR; noise from WHAM (tr only) convolved with K ~ U(1, 5) noise RIRs and summed, mixed at SNR ~ U(5, 40) dB.

• Scene-coherent. Source and noise RIRs come from the same GRAM scene (labhamlet/BinauralRIRs scene JSON), i.e. the same simulated room.
• Train houses only. Only the 70 GRAM train houses are used; references to the 15 held-out test houses: 0.
• 16 kHz (GRAM renders 32 kHz), 2-channel PCM16 FLAC, exactly 160,000 samples.
• Clips whose peak exceeded 0.99 were rescaled with one gain for both channels (53.1 % of clips); inter-channel level differences are preserved.
• Seed 20260915. Every random draw is fixed in a per-clip manifest, and the corpus re-renders bit-identically from it.

Layout

flac/NNNN.tar holds NNNN/<youtube_id>.flac. Tars are uncompressed: FLAC is already a lossless codec and gzip/zstd gained 0.2 % on a measured sample.

for t in flac/*.tar; do tar xf "$t"; done   # -> 0000/ ... 0063/

Many ids begin with --; that is harmless when extracting, but pass files to the tar CLI as ./--id.flac if you ever re-pack by name.

Access

Downloads are gated: request access on this page and it is granted manually.

Terms

Derived from third-party data; use is subject to their terms. AudioSet audio is YouTube content whose rights belong to its uploaders (the AudioSet labels are CC BY 4.0). WHAM! noise is CC BY-NC 4.0. labhamlet/BinauralRIRs is MIT. Research use only.