Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 81, in _split_generators
                  first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 32, in _get_pipeline_from_tar
                  fs: fsspec.AbstractFileSystem = fsspec.filesystem("memory")
                                                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 302, in filesystem
                  cls = get_filesystem_class(protocol)
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 239, in get_filesystem_class
                  raise ValueError(f"Protocol not known: {protocol}")
              ValueError: Protocol not known: memory
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 71, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

callcentre-noise48k

A call-centre-weighted background-noise corpus at 48 kHz mono, for noise augmentation in speech-enhancement and speech-restoration training.

20,946 chunks · ≈28.7 h · 9.92 GB · 48 kHz mono WAV (PCM_16)

This is a derived corpus: five public noise/event datasets, decoded to 48 kHz mono, cut into ≤10 s chunks, and re-labelled into one flat category scheme chosen for telephony-domain relevance. It contains no speech content — see Deliberate omissions.

Files

file size
archive/callcentre-noise48k.tar.part-aa 4.29 GB
archive/callcentre-noise48k.tar.part-ab 4.29 GB
archive/callcentre-noise48k.tar.part-ac 1.33 GB
noise_filelist.tsv 1.1 MB

The tar is split only to stay under per-file limits — the three parts are a plain byte split of one uncompressed tar, so they must be concatenated in order:

huggingface-cli download Scicom-intl/callcentre-noise48k --repo-type dataset --local-dir .
mkdir -p /data/noise48k
cat archive/callcentre-noise48k.tar.part-a? | tar xf - -C /data/noise48k

# the TSV stores the absolute paths from the machine that built it — repoint them
sed -i 's#/root/noise48k#/data/noise48k#g' /data/noise48k/noise_filelist.tsv

noise_filelist.tsv is also included inside the tar, so the extracted tree is self-contained. The Dataset Viewer is disabled: this is a tar of WAVs, not parquet.

noise_filelist.tsv

Tab-separated, no header, 20,946 rows, three columns:

/root/noise48k/musan/noise-free-sound-0010_000.wav	office	musan
path	category	source

Consumers pick a chunk by weighting category for the acoustic scene they want to simulate; source is kept so a subset can be filtered by upstream licence (see Licensing — this matters).

Composition

Chunks per category × source:

category musan demand esc50 fsd50k audioset total
street 2880 1564 600 5044
office 2515 1440 120 36 129 4240
ambient 1520 1589 267 3376
cafe 1440 40 263 19 1762
car 1440 160 1600
domestic 1440 1440
hold_music 996 99 105 1200
babble 855 52 907
keyboard 80 690 98 868
phone 191 79 270
hvac 80 103 56 239
total 3511 8640 2000 5390 1405 20946

office and street dominate by design — they are the two scenes a call-centre caller is most often in. MUSAN's music partition was downsampled to 996 chunks so hold-music would not swamp the corpus (it is by far the largest raw partition). FSD50K and AudioSet were class-filtered by label keyword down to call-centre-relevant events; unmatched labels were dropped rather than binned as "other", which is why their category spreads are narrow.

Processing

Every chunk went through the same path:

  1. decode → 48 kHz mono float32 (soxr resampling)
  2. cut into fixed 10 s frames, discarding any tail shorter than 2 s
  3. drop near-silent frames (peak < 1e-4)
  4. peak-limit to 0.9 only if the frame clipped (peak > 1) — otherwise the original level is preserved, so relative loudness across a source survives and consumers can set their own SNR
  5. write WAV PCM_16

Chunks are therefore variable length between 2 s and 10 s, averaging ≈4.9 s — not a uniform 10 s. Compute duration from the files, not from rows × 10 s (that would overstate the corpus at 58 h). The ≈28.7 h figure above is derived from the payload size: 9.91 GB at 48 kHz mono PCM_16 is 96 kB/s.

Deliberate omissions

  • No speech, and no multi-speaker babble beds. The babble category here is crowd/chatter ambience (from FSD50K and AudioSet crowd labels) — a room tone, not intelligible talkers. If you need overlapping-speaker interference, mix it yourself from a speech corpus.
  • No RIRs. These chunks are dry recordings; reverberation is not baked in.
  • Nothing scraped. No YouTube downloading was performed; the AudioSet portion came from an existing Hugging Face audio mirror.

Licensing

Read this before any commercial use. This corpus is a derivative of five upstream datasets whose licences differ, and it is not uniformly permissive. The repo licence is therefore other; the effective terms are the union of:

source chunks upstream licence commercial use
MUSAN (OpenSLR 17) 3511 CC BY 4.0 yes, with attribution
DEMAND (Zenodo) 8640 CC BY-SA 3.0 yes, share-alike
ESC-50 2000 CC BY-NC 3.0 no — non-commercial only
FSD50K (Zenodo) 5390 CC BY 4.0 overall, but per-clip CC0 / CC BY / CC BY-NC / Sampling+ mixed — check per clip
AudioSet (via HF mirror) 1405 labels CC BY 4.0; the audio originates from YouTube and is not covered by that licence unclear — treat as restricted

Practical consequences:

  • As distributed, treat the whole corpus as research / non-commercial.
  • For a commercially clean subset, filter noise_filelist.tsv on the source column to musan and demand only (12,151 of 20,946 chunks) and observe CC BY-SA 3.0's share-alike obligation for the DEMAND portion.
  • The source column exists precisely so this filtering is a one-liner:
awk -F'\t' '$3=="musan" || $3=="demand"' noise_filelist.tsv > noise_commercial.tsv

Please cite the upstream datasets, not just this repo — MUSAN, DEMAND, ESC-50, FSD50K and AudioSet each have a paper.

Downloads last month
87