🔊 MMAUD extracted background noise

6,588 NumPy files · 12 original groups · 5.057 GB

Background-noise arrays extracted for MMAUD and supplied by the dataset maintainer. The original np_data_align folder structure, filenames and file bytes are preserved.

Browse / download · Original Google Drive folder · Source manifest · SHA-256 checksums

Array format

Each .npy file stores a two-dimensional float32 array with 4 columns. Nonempty arrays have 48,000 rows, giving shape (48000, 4). Load them with numpy.load(..., allow_pickle=False).

The source contains 4 empty arrays with shape (0, 4): 10/0.npy, 11/0.npy, 12/0.npy, b9/0.npy. These files are kept unchanged. Skip them when constructing batches.

Folder names are retained exactly, including 10, 11 and 12. Filenames are numeric source indices; the mirror does not rename or resample the arrays.

Files by group

Group	NPY files	Size (decimal MB)	Empty arrays
b1	1,183	908.695	0
b2	898	689.779	0
b3	1,063	816.520	0
b4	1,621	1245.135	0
b5	978	751.229	0
b6	79	60.682	0
b7	45	34.566	0
b8	48	36.870	0
b9	174	132.886	1
10	168	128.278	1
11	192	146.713	1
12	139	106.002	1
Total	6,588	5057.355	4

Load one file

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(
    repo_id="Yuan-Shenghai/MMAUD",
    repo_type="dataset",
    filename="background_noise/np_data_align/b1/5.npy",
)
array = np.load(path, allow_pickle=False)
print(array.shape, array.dtype)
  

Download a group or the full collection

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Yuan-Shenghai/MMAUD",
    repo_type="dataset",
    local_dir="MMAUD",
    allow_patterns=["background_noise/np_data_align/b1/*"],
)
  

Use allow_patterns=["background_noise/*"] to download all 12 groups, this guide, the reading example and the checksum list. From the downloaded MMAUD directory, run python background_noise/load_example.py background_noise/np_data_align/b1/5.npy to inspect a local file.

Provenance and verification

Published on 28 October 2026 from the maintainer-provided public Drive folder. Every original file was checked against Google's CRC32C response, parsed as NumPy data and assigned a SHA-256 digest. Every mirrored array was then downloaded anonymously from Hugging Face in full and matched by size and SHA-256. The source manifest records every Drive file ID and its repository path.

This addition is listed separately from the original V1/V2/V3 collection and its supplementary/noise.bin file. It does not replace any earlier recording.

License and citation

Distributed with MMAUD under CC BY-NC-SA 4.0. Cite the MMAUD ICRA 2024 paper; the dataset BibTeX is available separately.