BAS4R: A Multi-Condition Bangla Speech Dataset for Gender-Aware Real and Fake Voice Analysis (doi:10.7910/DVN/YVWOAL)

View:

Part 1: Document Description
Part 2: Study Description
Part 5: Other Study-Related Materials
Entire Codebook

Document Description

Citation

Title:

BAS4R: A Multi-Condition Bangla Speech Dataset for Gender-Aware Real and Fake Voice Analysis

Identification Number:

doi:10.7910/DVN/YVWOAL

Distributor:

Harvard Dataverse

Date of Distribution:

2026-02-11

Version:

1

Bibliographic Citation:

Al Arian Ahmad; Turzo, Nakib; Durjoy Kumar Dutta; Wadud, Abdul Wadud, 2026, "BAS4R: A Multi-Condition Bangla Speech Dataset for Gender-Aware Real and Fake Voice Analysis", https://doi.org/10.7910/DVN/YVWOAL, Harvard Dataverse, V1

Study Description

Citation

Title:

BAS4R: A Multi-Condition Bangla Speech Dataset for Gender-Aware Real and Fake Voice Analysis

Identification Number:

doi:10.7910/DVN/YVWOAL

Authoring Entity:

Al Arian Ahmad (https://ror.org/01vxg3438)

Turzo, Nakib (https://ror.org/01vxg3438)

Durjoy Kumar Dutta (https://ror.org/01vxg3438)

Wadud, Abdul Wadud (https://ror.org/01vxg3438)

Distributor:

Harvard Dataverse

Access Authority:

Al Arian Ahmad

Depositor:

Al Arian Ahmad

Date of Deposit:

2026-02-11

Holdings Information:

https://doi.org/10.7910/DVN/YVWOAL

Study Scope

Keywords:

Computer and Information Science, Engineering

Abstract:

BAS4R is a structured and large-scale Bangla speech dataset developed to support research in replay attack detection and audio spoofing analysis within voice biometric systems. The dataset contains both authentic (real) and systematically manipulated (spoofed) speech recordings collected under controlled and realistic acoustic conditions. The complete dataset comprises 143.88 hours of audio recordings, totaling 120,125 audio files, organized into five major categories: Channel-based: 28,830 files (34.65 hours) Signal Processing-based: 28,830 files (34.53 hours) Effect-based: 28,830 files (34.48 hours) Replay-based: 28,830 files (34.48 hours) Real Data: 4,805 files (5.75 hours) Speech samples were collected from 100 native Bangla speakers (50 male and 50 female) aged 20–26 years, ensuring balanced gender representation and demographic consistency. All recordings were captured in controlled environments and stored in high-quality digital audio format. The dataset follows a structured hierarchical organization separating real and spoofed samples by category and attack condition, facilitating reproducible research. The spoofed data were generated using real signal processing techniques, channel transmission effects, environmental distortions, and replay setups. BAS4R is suitable for research in anti-spoofing systems, speaker verification robustness evaluation, replay attack detection, and deep learning–based audio classification.

Methodology and Processing

Sources Statement

Data Access

Notes:

<a href="http://creativecommons.org/publicdomain/zero/1.0">CC0 1.0</a>

Other Study Description Materials

Other Study-Related Materials

Label:

Channel-based.zip

Notes:

application/zip

Other Study-Related Materials

Label:

Effect-based.zip

Notes:

application/zip

Other Study-Related Materials

Label:

Real Data.zip

Notes:

application/zip

Other Study-Related Materials

Label:

Replay-based.zip

Notes:

application/zip

Other Study-Related Materials

Label:

Signal Processing.zip

Notes:

application/zip