vdivyasharma/IndicSynth

IndicSynth: Indian Multilingual Audio Deepfake Detection & Anti-Spoofing Dataset

A Large-Scale Multilingual Synthetic Speech Dataset for Low-Resource Indian Languages to facilitate audio deepfake detection and anti-spoofing research

🏆 Outstanding Paper Award, ACL 2025

🧠 Overview

IndicSynth is a novel multilingual synthetic speech dataset designed to advance multilingual audio deepfake detection (ADD) and anti-spoofing research. It covers 12 low-resource Indian languages and provides both mimicry and diversity subsets.

• —4,000+ hours of synthetic audio
• —12 languages: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Sanskrit, Tamil, Telugu, Urdu
• —Useful for:

1. 1.Multilingual audio deepfake detection (ADD) research or mitigating linguistic biases in audio deepfake detection systems
2. 2.Enhancing the robustness of speaker verification (SV) systems against spoofing (impersonation) attacks and developing anti-spoofing solutions
3. 3.Cross-lingual or gender bias studies in speech synthesis and recognition systems

📂 Dataset Structure

Each language folder contains:

<pre> IndicSynth/ ├── Bengali/ │ ├── audio/ # All .wav files (synthetic clips) │ └── metadata.csv # Metadata for all synthetic clips ├── Gujarati/ │ ├── audio/ │ └── metadata.csv </pre> Each 'metadata.csv' includes:

• —Generative Model (xtts_v2 / vits / freevc24)
• —Speaker IDs
• —Gender
• —Transcript (if applicable)
• —File path to synthetic audio

Bona fide audios: The bona fide source and target speech samples referenced in IndicSynth metadata are drawn from the IndicSUPERB dataset. Please refer to the official repository for details: https://github.com/AI4Bharat/indicSUPERB.

📝 Note on Transcripts in Metadata: The transcripts included in the metadata.csv files represent the intended text prompts used during synthetic speech generation via TTS models. We provide these transcripts to enable future explorations, but do not guarantee perfect alignment with the generated audio. If you intend to use IndicSynth for speech-to-text or similar tasks, we strongly recommend conducting careful human evaluation with proficient native speakers of the respective languages.

⚙️ IndicSynth Generation?

Synthetic data was generated using:

• —Mimicry subset: For anti-spoofing research
• —Diversity subset: Contains diverse set of realistic synthetic voices for multilingual audio deepfake detection research

For more details, please see the Table 1 and Section 3 of our paper: https://aclanthology.org/2025.acl-long.1070.pdf

📦 Access the Dataset

You can load data in a specific target language using the following code:

import os
import soundfile as sf
from datasets import load_dataset
from tqdm import tqdm
import pandas as pd

language = "Hindi" # Specify the target language here

# Load Dataset
dataset = load_dataset("vdivyasharma/IndicSynth", name=language, split="train") 

# Create target directory structure
output_dir = language
audio_dir = os.path.join(output_dir, "audio")
os.makedirs(audio_dir, exist_ok=True)

# Store metadata rows here
metadata_rows = []

# Loop through dataset and save each clip
for example in tqdm(dataset):
    audio_array = example["audio"]["array"]
    sampling_rate = example["audio"]["sampling_rate"]
    
    # Get filename
    original_name = example.get("file") or example.get("path") or example["audio"]["path"].split("/")[-1]
    
    # Save audio to audio/ subfolder
    audio_path = os.path.join("audio", original_name)  # relative path for metadata
    sf.write(os.path.join(output_dir, audio_path), audio_array, sampling_rate)
    
    # Store metadata row
    row = {k: v for k, v in example.items() if k != "audio"}
    row["file_name"] = audio_path
    metadata_rows.append(row)

# Save metadata to CSV
df = pd.DataFrame(metadata_rows)
df.to_csv(os.path.join(output_dir, "metadata.csv"), index=False)

License

IndicSynth is released under the CC BY-NC 4.0 License. It is intended for non-commercial, academic research only.

Citation

If you use IndicSynth, please cite the following papers: <pre>@inproceedings{sharma-etal-2025-indicsynth, title = "{I}ndic{S}ynth: A Large-Scale Multilingual Synthetic Speech Dataset for Low-Resource {I}ndian Languages", author = "Sharma, Divya V and Ekbote, Vijval and Gupta, Anubha", editor = "Che, Wanxiang and Nabende, Joyce and Shutova, Ekaterina and Pilehvar, Mohammad Taher", booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", month = jul, year = "2025", address = "Vienna, Austria", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2025.acl-long.1070/", pages = "22037--22060", ISBN = "979-8-89176-251-0" } </pre> <pre>@article{IndicSuperb, author = {Javed, Tahir and Bhogale, Kaushal and Raman, Abhigyan and Kumar, Pratyush and Kunchukuttan, Anoop and Khapra, Mitesh}, year = {2023}, month = {06}, pages = {12942-12950}, title = {IndicSUPERB: A Speech Processing Universal Performance Benchmark for Indian Languages}, volume = {37}, journal = {Proceedings of the AAAI Conference on Artificial Intelligence}, doi = {10.1609/aaai.v37i11.26521} } </pre> ---

💬 Contact

For questions or feedback, please feel free to reach out at divyas@iiitd.ac.in.

🙏 Acknowledgments

• —🌍 ACL Diversity & Inclusion Subsidy for enabling in-person presentation at ACL 2025
• —🤝 HuggingFace for dataset hosting support
• —🎓SBILab and Infosys Centre for Artificial Intelligence (CAI) at IIIT-Delhi for their support