Dataset Card for fed-urbansound8k

Dataset Description

This dataset is a federated, non-IID repartitioning of danavery/urbansound8K, a Hugging Face version of UrbanSound8K.

The source dataset contains labeled urban sound excerpts of up to 4 seconds from 10 urban sound classes. This derived dataset keeps the original audio examples and metadata, but reorganizes them into 50 simulated federated clients for experiments in client-partitioned audio classification.

Each row includes a clientID column identifying the simulated client/silo to which the example belongs.

Source Dataset

• Source dataset: danavery/urbansound8K
• Source split used: train
• Source task: urban sound classification
• Source size: 8,732 labeled audio excerpts
• Source classes:
  • air_conditioner
  • car_horn
  • children_playing
  • dog_bark
  • drilling
  • engine_idling
  • gun_shot
  • jackhammer
  • siren
  • street_music

The original UrbanSound8K examples are excerpts from field recordings uploaded to Freesound.

Dataset Construction

This dataset was generated by the accompanying processing script using the following procedure:

1. Load danavery/urbansound8K from Hugging Face.
2. Recover or preserve the fsID field, which identifies the original Freesound recording group.
3. Group samples by fsID so that all excerpts from the same source recording remain assigned to the same client.
4. Build per-fsID label-distribution tables.
5. Create 50 simulated clients.
6. Sample client label preferences from a Dirichlet distribution.
7. Sample client size targets from a log-normal distribution.
8. Greedily assign whole fsID groups to clients using the per-group label distribution, client label preferences, and client size targets.
9. Split each client locally into train/test subsets.
10. Concatenate all client-local train subsets into the global train split and all client-local test subsets into the global test split.
11. Preserve the audio column as a Hugging Face Audio feature.

Federated Partitioning Details

The partitioning is intentionally non-IID. Heterogeneity is introduced through:

• Source-group preservation: all samples with the same fsID are kept on the same client.
• Label skew: clients have different class preferences sampled from a Dirichlet distribution.
• Size imbalance: client sizes are sampled from a log-normal distribution.
• Minimum group constraints: the assignment process attempts to give each client at least a minimum number of fsID groups when possible.

Recommended reproducibility parameters from the generation script:

Parameter	Value
NUM_CLIENTS	50
DIRICHLET_ALPHA	1.0
SIZE_IMBALANCE_SIGMA	0.3
MIN_CLIENT_SIZE	20
MIN_FSID_PER_CLIENT	3
TEST_RATE	0.1
SEED	42

Splits

The dataset contains two global splits:

Split	Description
train	Union of all client-local training subsets
test	Union of all client-local test subsets

Both splits contain examples from the simulated clients and include the clientID column. The global test split is not a central random split; it is the concatenation of each client's local held-out test set.

Schema

The exact schema follows the upstream UrbanSound8K Hugging Face dataset, with fold removed and clientID added.

Expected columns include:

Column	Type	Description
audio	Audio	Audio waveform feature preserved from the source dataset
slice_file_name	string	Original audio slice filename, when present in the source data
fsID	integer	Freesound source recording identifier
start	float	Start time of the excerpt, when available
end	float	End time of the excerpt, when available
salience	integer	Foreground/background salience metadata, when available
classID	integer/class label	Numeric urban sound class identifier
class	string/class label	Human-readable urban sound class
clientID	integer	Simulated federated client identifier in [0, 49]

How to Use

from datasets import load_dataset

ds = load_dataset("flwrlabs/fed-urbansound8k")

print(ds)
print(ds["train"][0])

To inspect one client's local partition:

from flwr_datasets import FederatedDataset
from flwr_datasets.partitioner import NaturalIdPartitioner

fds = FederatedDataset(
    dataset="flwrlabs/fed-urbansound8k",
    partitioners={"train": NaturalIdPartitioner(partition_by="clientID")},
)
partition = fds.load_partition(partition_id=0)

For audio classification, use the appropriate label column from the source schema, typically classID or class.

Intended Uses

This dataset is intended for research and benchmarking in:

• federated learning
• non-IID audio classification
• client distribution shift
• source-grouped client partitioning
• robustness across heterogeneous audio clients
• personalized or clustered federated learning

Out-of-Scope Uses

This dataset should not be used as a substitute for newly collected real-world client audio data. The clients are simulated partitions of a public benchmark and may not reflect all operational constraints, privacy conditions, or distribution shifts found in deployed federated audio systems.

Limitations

• This is a derived repartitioning of UrbanSound8K, not a newly collected dataset.
• Client identities are simulated and do not correspond to real users, devices, or institutions.
• Because fsID groups are kept intact, the final client sizes and label proportions may only approximately match the sampled targets.
• Audio sampling rates, channel counts, and file properties follow the upstream dataset and may vary across examples.

Citation

Please cite the original UrbanSound8K work when using this dataset. If you're using this dataset with Flower Datasets, you can cite Flower.

@inproceedings{salamon2014dataset,
  title={A dataset and taxonomy for urban sound research},
  author={Salamon, Justin and Jacoby, Christopher and Bello, Juan Pablo},
  booktitle={Proceedings of the 22nd ACM international conference on Multimedia},
  pages={1041--1044},
  year={2014}
}

@article{DBLP:journals/corr/abs-2007-14390,
  author       = {Daniel J. Beutel and
                  Taner Topal and
                  Akhil Mathur and
                  Xinchi Qiu and
                  Titouan Parcollet and
                  Nicholas D. Lane},
  title        = {Flower: {A} Friendly Federated Learning Research Framework},
  journal      = {CoRR},
  volume       = {abs/2007.14390},
  year         = {2020},
  url          = {https://arxiv.org/abs/2007.14390},
  eprinttype    = {arXiv},
  eprint       = {2007.14390},
  timestamp    = {Mon, 03 Aug 2020 14:32:13 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2007-14390.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Dataset Card Contact

In case of any doubts, please contact Flower Labs.