IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages

Overview

INDICVOICES is a dataset of natural and spontaneous speech containing a total of 16347 hours of read (8%), extempore (76%) and conversational (15%) audio from 25K speakers covering 308 Indian districts and 22 languages. Of these 16347 hours, 7256 hours have already been transcribed. Through this paper, we share our journey of capturing the cultural, linguistic and demographic diversity of India to create a one-of-its-kind inclusive and representative dataset. More specifically, we share an open-source blueprint for data collection at scale comprising of standardised protocols, centralised tools, a repository of engaging questions, prompts and conversation scenarios spanning multiple domains and topics of interest, quality control mechanisms, comprehensive transcription guidelines and transcription tools. We hope that this open source blueprint will serve as a comprehensive starter kit for data collection efforts in other multilingual regions of the world. Using INDICVOICES, we build IndicASR, the first ASR model to support all the 22 languages listed in the 8th schedule of the Constitution of India.

This work is funded by Bhashini, MeitY and Nilekani Philanthropies

Usage

The datasets library enables you to load and preprocess the dataset directly in Python. Ensure you have an active HuggingFace access token (obtainable from Hugging Face settings) before proceeding.

To load the dataset, run:

from datasets import load_dataset
# Load the dataset from the HuggingFace Hub
dataset = load_dataset("ai4bharat/IndicVoices","assamese",split="valid")
# Check the dataset structure
print(dataset)

You can also stream the dataset by enabling the streaming=True flag:

from datasets import load_dataset
dataset = load_dataset("ai4bharat/IndicVoices","assamese",split="valid", streaming=True)
print(next(iter(dataset)))

Citation

If you use IndicVoices in your work, please cite us:

@inproceedings{DBLP:conf/acl/JavedNGJBMSAFPR24,
  author       = {Tahir Javed and
                  Janki Nawale and
                  Eldho Ittan George and
                  Sakshi Joshi and
                  Kaushal Santosh Bhogale and
                  Deovrat Mehendale and
                  Ishvinder Virender Sethi and
                  Aparna Ananthanarayanan and
                  Hafsah Faquih and
                  Pratiti Palit and
                  Sneha Ravishankar and
                  Saranya Sukumaran and
                  Tripura Panchagnula and
                  Sunjay Murali and
                  Kunal Sharad Gandhi and
                  Ambujavalli R and
                  Manickam K. M and
                  C. Venkata Vaijayanthi and
                  Krishnan Srinivasa Raghavan Karunganni and
                  Pratyush Kumar and
                  Mitesh M. Khapra},
  title        = {IndicVoices: Towards building an Inclusive Multilingual Speech Dataset
                  for Indian Languages},
  booktitle    = {{ACL} (Findings)},
  pages        = {10740--10782},
  publisher    = {Association for Computational Linguistics},
  year         = {2024}
}

License

This dataset is released under the CC BY 4.0.

Contact

For any questions or feedback, please contact:

• Tahir Javed ([email protected])
• Mitesh M. Khapra ([email protected])