Dataset Card for SpARZ

Dataset Summary

The SpARZ dataset is designed for Audio Visual Zooming tasks. It was introduced and accepted at IECON 26 for the paper titled: "RDSE: Real-Time Directional Speech Enhancement Framework for Dual-Mic Commodity Smartphones". The dataset features complex room acoustics, varying reverberation times (RT60), and different interference angles and Signal-to-Noise Ratios (SNRs) to simulate realistic spatial audio environments.

Paper Evaluation Subset: While this repository contains the full, comprehensively varied dataset, the results reported in the IECON 26 paper were evaluated on a specific subset referred to in the paper as SpARZ 100. This corresponds to the simulated room with dimensions 4.9m x 4.9m x 4.9m at two specific reverberation times: RT60 = 0.0 and RT60 = 0.5. If you wish to reproduce or compare against the paper's exact results on the SpARZ 100 subset, you should use the folders for this specific room setting (e.g., room_4.9x4.9x4.9m__rt60_0.0__ang_30__snr_5 and room_4.9x4.9x4.9m__rt60_0.5__ang_30__snr_5).

Supported Tasks and Leaderboards

• audio-visual-zooming
• speech-enhancement
• source-separation

Languages

English (en)

Dataset Structure

How to Use

The audio data is distributed as a single SpARZ.tar archive to preserve directory structures and speed up downloading. To use the dataset, simply download the .tar file and extract it:

tar -xf SpARZ.tar

Data Instances

Each sample in the dataset is structured within a sub-folder that defines the acoustic parameters (e.g., room dimensions, RT60, interference angle, SNR). Inside each sample folder (e.g., sample_00), there is a meta.json containing the metadata for the audio sample.

Example meta.json:

{
  "sample_id": 0,
  "cell_id": 0,
  "room_dims": [3.0, 4.0, 3.0],
  "rt60": 0.0,
  "target_angle_deg": 90,
  "interferer_angle_deg": 0,
  "sir_db": 0.0,
  "snr_db": 0,
  "target_audio_path": "datasets/LibriSpeech/train-clean-100/302/123504/302-123504-0011.flac",
  "interferer_audio_path": "datasets/musan/speech/us-gov/speech-us-gov-0210.wav",
  "clip_duration_s": 5.0,
  "sample_rate": 16000
}

Data Fields

• sample_id: Unique identifier for the sample.
• cell_id: Spatial cell identifier.
• room_dims: Dimensions of the simulated room [x, y, z] in meters.
• rt60: Reverberation time in seconds.
• target_angle_deg: Angle of the target source in degrees.
• interferer_angle_deg: Angle of the interfering source in degrees.
• sir_db: Signal-to-Interference Ratio in dB.
• snr_db: Signal-to-Noise Ratio in dB.
• target_audio_path: Path to the original target audio (e.g., LibriSpeech).
• interferer_audio_path: Path to the original interferer audio (e.g., MUSAN).
• clip_duration_s: Duration of the audio clip in seconds.
• sample_rate: Sampling rate of the audio in Hz.

Data Splits

The dataset is provided as a single, unified collection of audio samples. There are no predefined Train, Validation, or Test splits. Researchers can split the dataset based on their specific experimental requirements.

Dataset Creation

Curation Rationale

The SpARZ dataset was developed to advance research in audio visual zooming and real-time directional speech enhancement by providing highly realistic, parametrically varied acoustic environments.

Source Data

Initial Data Collection and Normalization

The target speech is sourced from LibriSpeech and the interferer noise/speech is sourced from MUSAN. These are mixed synthetically using spatial acoustic simulators to generate the final audio.

Additional Information

Dataset Curators

This dataset was developed as part of a collaborative research effort by the authors of the RDSE framework (Chinmay Deo, Danish Mahajan, Soudeep Ghoshal, and Himanshu Buckchash).

Licensing Information

Creative Commons Attribution 4.0 International (CC-BY-4.0)

Citation Information

If you use this dataset, please cite the corresponding IECON 26 paper:

@misc{deo2026rdserealtimedirectionalspeech,
      title={RDSE: Real-Time Directional Speech Enhancement Framework for Dual-Mic Commodity Smartphones}, 
      author={Chinmay Deo and Danish Mahajan and Soudeep Ghoshal and Himanshu Buckchash},
      year={2026},
      eprint={2610.12476},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2610.12476}, 
}