Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BanglaFake: Constructing and Evaluating a Specialized Bengali Deepfake Audio Dataset

Istiaq Ahmed Fahad, Kamruzzaman Asif, Sifat Sikder

TL;DR

This work tackles the scarcity of Bengali deepfake audio data by introducing BanglaFake, a benchmark containing 12,260 real and 13,260 deepfake utterances generated with a VITS-based TTS system. The dataset leverages phonetically balanced Bengali text and combines SUST TTS corpus audio with Mozilla Common Voice samples, enhancing realism for detection tasks. Qualitative MOS (naturalness 3.40) and intelligibility (4.01), along with MFCC-based t-SNE visualizations indicating real–fake overlap, demonstrate the dataset's challenging nature for detectors. Public availability on HuggingFace enables researchers to train and benchmark Bengali deepfake detectors, advancing AI-forensics in low-resource languages.

Abstract

Deepfake audio detection is challenging for low-resource languages like Bengali due to limited datasets and subtle acoustic features. To address this, we introduce BangalFake, a Bengali Deepfake Audio Dataset with 12,260 real and 13,260 deepfake utterances. Synthetic speech is generated using SOTA Text-to-Speech (TTS) models, ensuring high naturalness and quality. We evaluate the dataset through both qualitative and quantitative analyses. Mean Opinion Score (MOS) from 30 native speakers shows Robust-MOS of 3.40 (naturalness) and 4.01 (intelligibility). t-SNE visualization of MFCCs highlights real vs. fake differentiation challenges. This dataset serves as a crucial resource for advancing deepfake detection in Bengali, addressing the limitations of low-resource language research.

BanglaFake: Constructing and Evaluating a Specialized Bengali Deepfake Audio Dataset

TL;DR

This work tackles the scarcity of Bengali deepfake audio data by introducing BanglaFake, a benchmark containing 12,260 real and 13,260 deepfake utterances generated with a VITS-based TTS system. The dataset leverages phonetically balanced Bengali text and combines SUST TTS corpus audio with Mozilla Common Voice samples, enhancing realism for detection tasks. Qualitative MOS (naturalness 3.40) and intelligibility (4.01), along with MFCC-based t-SNE visualizations indicating real–fake overlap, demonstrate the dataset's challenging nature for detectors. Public availability on HuggingFace enables researchers to train and benchmark Bengali deepfake detectors, advancing AI-forensics in low-resource languages.

Abstract

Deepfake audio detection is challenging for low-resource languages like Bengali due to limited datasets and subtle acoustic features. To address this, we introduce BangalFake, a Bengali Deepfake Audio Dataset with 12,260 real and 13,260 deepfake utterances. Synthetic speech is generated using SOTA Text-to-Speech (TTS) models, ensuring high naturalness and quality. We evaluate the dataset through both qualitative and quantitative analyses. Mean Opinion Score (MOS) from 30 native speakers shows Robust-MOS of 3.40 (naturalness) and 4.01 (intelligibility). t-SNE visualization of MFCCs highlights real vs. fake differentiation challenges. This dataset serves as a crucial resource for advancing deepfake detection in Bengali, addressing the limitations of low-resource language research.
Paper Structure (21 sections, 2 figures, 2 tables)

This paper contains 21 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Related Work
  4. Deepfake Detection Models
  5. Deepfake Detection Datasets
  6. Low-Resource Language TTS and Deepfake Generation
  7. Dataset Description
  8. Dataset Overview
  9. Data Composition
  10. Dataset Structure
  11. Dataset Usage
  12. Dataset Availability
  13. Methodology
  14. Model Training
  15. Fake Audio Generation
  16. ...and 6 more sections

Figures (2)

  • Figure 1: Overview of the Bangla Fake Speech Dataset Generation Pipeline. The left side illustrates the process of extracting latent speech representations from the SUST TTS corpus using a posterior encoder, normalizing flow, and stochastic duration predictor. The right side depicts the synthesis process, where text from the Bangla dataset is encoded and converted into speech-like representations using a trained model. The final synthesized speech dataset follows the LJSpeech format
  • Figure 2: t-SNE visualization of MFCC features for real and deepfake audio samples. Real audio samples are represented in blue circles, while deepfake audio samples are shown in red squares. The overlap between the two classes highlights the challenge of distinguishing fake audio from real audio.