Project Vaani, by IISc, Bangalore and ARTPARK, is capturing the true diversity of India’s spoken languages to propel language AI technologies and content for an inclusive Digital India.
We expect to create data corpora of over 150,000 hours of speech, part of which will be transcribed in local scripts, while ensuring linguistic, educational, urban-rural, age, and gender diversity (among other potential diversity characteristics). These diligently collected and curated datasets of natural speech and text from about 1 million people across all 773 districts of India will be open-sourced. The current version of the data is open-sourced here. Going forward, we hope to open source through platforms like Bhashini (under the National Language Translation Mission, MeiTY).
This will boost the development of technologies such as automatic speech recognition (ASR), speech to speech translation (SST), and natural language understanding (NLU) that reflect the ground realities of how Indians speak.
Google is funding the Project Vaani.
VAANI dataset covers 59 languages collected form 80 districts across India. It contains 83174 speakers utterance saved in 8834039 files with duration of 14434 hours. Out this data, we have transcribed 349 hours of data spanning almost equally across the 80 districts. In the data collection process, we have used 129028 images to collect the audios.
To learn more, please click here
How to use the data
You can load the dataset using the load_dataset function from the datasets library. When loading the dataset, you can specify which subset (formatted as {State}_{District}) you would like to access. Please refer to the dataset card to explore the available subsets. Additionally, you will need to use an access token, which you can obtain here, to access the dataset.
from datasets import load_dataset
dataset_subset = load_dataset("ARTPARK-IISc/VAANI", "Goa_NorthSouthGoa",token=<access_token>)
Contact Us
We are always open to new ideas and collaborations. Feel free to reach out to us at vaanicontact@gmail.com
- Downloads last month
- 826