Vaani
Vaani is a multilingual multimodal dataset representative of Indian linguistic diversity. It contains hours of spontaneous audio recordings accompanied by images and transcripts covering 86 languages, collected from over 112,000 speakers in 120 districts. This dataset is designed to improve inclusive and multilingual AI technologies.
Approximately 21,500 hours of audio, 835 hours of text transcripts, 210,000 images, audio, text, and various image formats
CC-BY 4.0
Description
The dataset Vaani contains a vast collection of spontaneous audio recordings in 86 Indian languages, accompanied by images and text transcripts. These data represent the cultural, linguistic, and geographic diversity of India, from more than 112,000 speakers.
What is this dataset for?
- Training multilingual speech recognition models
- Develop multimodal AI systems combining audio, text, and image
- Promote research in the automatic processing of Indian languages
Can it be enriched or improved?
Yes, the dataset can be enriched by additional transcriptions, linguistic annotations, or extensions to more languages and dialects. Improving metadata and cleaning audio data is also possible.
🔎 In summary
🧠 Recommended for
- Speech-to-text researchers
- Multimodal AI
- Computational linguists
🔧 Compatible tools
- Kaldi
- Hugging Face
- ESPnet
- PyTorch
- TensorFlow
💡 Tip
Leverage existing transcripts to train ASR models adapted to Indian languages.
Frequently Asked Questions
What languages are covered by the Vaani dataset?
The dataset covers 86 Indian languages with a wide geographic and demographic diversity.
Is all audio transcribed?
No, about 835 hours are transcribed, the rest is raw audio.
Can I use this dataset for a commercial project?
Yes, licensed under CC-BY 4.0, with mandatory attribution.
