Datasets:

vdivyasharma
/

IndicSynth

license: cc-by-nc-4.0
configs:
  - config_name: Bengali
    data_files:
      - split: train
        path: Bengali/train-*
  - config_name: Gujarati
    data_files:
      - split: train
        path: Gujarati/train-*
dataset_info:
  - config_name: Bengali
    features:
      - name: audio
        dtype: audio
      - name: Generative Model
        dtype: string
      - name: Source Speaker_ID
        dtype: float64
      - name: Target Speaker ID
        dtype: int64
      - name: Gender
        dtype: string
      - name: Source Reference Audio
        dtype: string
      - name: Target Reference Audio
        dtype: string
      - name: TTS Transcript
        dtype: string
    splits:
      - name: train
        num_bytes: 26636977797.52
        num_examples: 83448
    download_size: 33644253977
    dataset_size: 26636977797.52
  - config_name: Gujarati
    features:
      - name: audio
        dtype: audio
      - name: Generative Model
        dtype: string
      - name: Source Speaker_ID
        dtype: float64
      - name: Target Speaker ID
        dtype: int64
      - name: Gender
        dtype: string
      - name: Source Reference Audio
        dtype: string
      - name: Target Reference Audio
        dtype: string
      - name: TTS Transcript
        dtype: string
    splits:
      - name: train
        num_bytes: 35729511328.708
        num_examples: 118778
    download_size: 50811268801
    dataset_size: 35729511328.708

IndicSynth

A Large-Scale Multilingual Synthetic Speech Dataset for Low-Resource Indian Languages

🏆 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 ADD, speaker verification (SV), and bias studies

📂 Dataset Structure

Each language folder contains:

  
IndicSynth/
├── Bengali/
│ ├── audio/ # All .wav files (synthetic clips)
│ └── metadata.csv # Metadata for all synthetic clips
├── Gujarati/
│ ├── audio/
│ └── metadata.csv

Each 'metadata.csv' includes:

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

📝 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:

Model	Type	Transcript	Fine-Tuned
xtts_v2	TTS	Yes	Yes (for 10 languages)
vits	TTS	Yes	No
freevc24	VC	No	No

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 a specific language using:

from datasets import load_dataset

ds = load_dataset("vdivyasharma/IndicSynth", name="Hindi", split="train")

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:

@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"
}

@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}
}

💬 Contact

For questions or feedback, please feel free to reach out at [email protected].

🙏 Acknowledgments

🌍 ACL Diversity & Inclusion Subsidy for enabling in-person presentation at ACL 2025

🤝 HuggingFace for dataset hosting support