Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains de-identified human speech recordings. Access is restricted to non-commercial research; you agree not to attempt to re-identify speakers or redistribute the audio.
Log in or Sign Up to review the conditions and access this dataset content.
Hiring Accent Speech (human voices)
A ~4.3-hour real-world speech dataset simulating job-interview scenarios across diverse accent profiles: 47 speakers from five accent groups — American, British, Chinese, Indian, and Nigerian English — recording the same scripted and unscripted hiring-interview prompts (282 clips, 16 kHz mono WAV). Because speakers read identical scripts, the corpus is designed for studying accent and fairness effects in hiring-style speech evaluation.
This dataset accompanies our research on implicit accent bias in audio language models — testing whether models judge speakers differently based on accent when content is held fixed. Code, analysis, and paper: https://github.com/arunasrivastava/ImplicitAccentBias.
Access, consent & privacy
- Gated. Access is granted for non-commercial research (see terms above).
- Consent. Only speakers who consented to recording release are included; 5 of the 52 study speakers are omitted (their de-identified ratings still appear in the code repo, but their audio is not released).
- Privacy. Speaker names are removed from filenames and metadata and replaced with stable IDs (
speaker_01, …) that match the analysis CSVs in the code repo. The audio itself is not de-identified (voices are inherently identifying; some unscripted clips contain spoken names) — hence the gating.
Recording protocol
Participants recorded on their own devices in a quiet, low-echo space (phones held 6–8 inches away), reading each scenario aloud and naturally as if in a real interview. Four scripted scenarios (personal introduction, personal commitment, financial product, client disagreement) — read verbatim so score differences isolate delivery — plus two unscripted prompt sets (a professional introduction and a recent disagreement). Exact scripts: https://github.com/arunasrivastava/ImplicitAccentBias/tree/main/data/human_hiring_corpus
Structure
<category>/<prefix> - <speaker_id>.wav # 16 kHz mono audio
metadata.csv # file_name, speaker_id, accent, demographics, ...
- Downloads last month
- 14