Oye12/A_Pilot_Speech_Corpus

A Pilot Speech Corpus for Studying Device and Environmental Variability in Voice Biometrics

This paper introduces a curated speech dataset developed to support research in speech enhancement and robust voice biometric verification.

The dataset addresses the critical challenge of voice variability across devices and environments, which can significantly reduce the reliability of speaker verification systems in real-world applications such as mobile banking.

A total of 480 recordings were collected from 12 participants with diverse demographic backgrounds, including individuals from Japan, Nigeria, the Ivory Coast, France, Germany, and Indonesia.

Structure Each participant contributed 40 utterances, recorded across multiple devices: Samsung A04s, OnePlus Nord, iPhone 15 Pro, and a USB condenser microphone on a MacBook, and under both indoor and outdoor conditions.

The recordings capture variations introduced by microphone quality, network transmission (in-call speech), and environmental noise.

All files are stored in WAV format with standardized sampling rates between 8–16 kHz, accompanied by metadata containing non-identifiable demographic attributes such as nationality, age, gender, and English proficiency.

*The dataset is designed for evaluating and benchmarking speech enhancement algorithms, including spectral subtraction, Wiener filtering, and adaptive filtering, as well as for studying their impact on downstream speaker identification and verification tasks.

*By providing a balanced collection of cross-device, cross-environment recordings, this dataset enables the development of more consistent, accurate, and secure voice biometric systems.

Contribution

• Benchmark resource for speech biometrics: The dataset captures cross-device, cross-environment variability that is underrepresented in existing corpora like VoxCeleb or LibriSpeech.
• Practical relevance: Collected with everyday devices (low- to high-end smartphones, in-call channels), reflecting real-world conditions faced in mobile banking authentication.
• Security applications: Supports robustness testing, spoofing resilience, and liveness detection research.
• Low-resource settings: Valuable for studies where large datasets are unavailable, providing a pilot-scale corpus for transfer learning and comparative evaluation.
• Reusability: Organized with speaker-wise folders and metadata files for easy integration into machine learning pipelines.

Data Format

• The main fields per entry in the dataset are: Participant ID, Nationality, English Proficiency, Age, Gender, Status.
• The data is available in both .csv and .JSON formats.

Below is an example of an entry from our dataset:

    {
        "participant_id":"ZVMKSJ",
        "nationality":"Indonesia",
        "english_proficiency":"Basic",
        "age":38.0,
        "gender":"Male",
        "status":"present"
    }

Usage and Licence Notes

We are releasing this dataset under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. To download the dataset, please use:

>>> from datasets import load_dataset
>>> dataset = load_dataset("Oye12/A_Pilot_Speech_Corpus")

Citation If you find our dataset relevant to your work, please cite our paper below

@inproceedings{oyewale_voicebiometric_2024,
  author    = {Oyewale, Oyebode Oluwatobi and Hossain, Md Delwar and Taenaka, Yuzo and Kadobayashi, Youki},
  title     = {Optimizing Voice Biometric Verification in Banking with Machine Learning for Speaker Identification},
  booktitle = {Proceedings of the 2024 IEEE 29th Asia Pacific Conference on Communications (APCC)},
  year      = {2024},
  address   = {Bali, Indonesia},
  pages     = {377--384},
  doi       = {10.1109/APCC62576.2024.10768085},
  url       = {https://doi.org/10.1109/APCC62576.2024.10768085}
}


@dataset{oyewale_2025_figshare,
  author       = {Oyebode Oluwatobi Oyewale},
  title        = {A Pilot Speech Corpus for Studying Device and Environmental Variability in Voice Biometrics},
  year         = {2025},
  publisher    = {Figshare},
  doi          = {10.6084/m9.figshare.30039037},
  url          = {https://doi.org/10.6084/m9.figshare.30039037}
}