VAANI: Multi-modal, Multi-lingual Dataset

About Dataset

The VAANI dataset captures spontaneous speech elicited via image prompts, encouraging rich and natural linguistic expression across diverse Indian languages. It is a high-quality, multimodal, and multilingual resource, curated from 120 districts across 22 Indian states. Each data point in VAANI comprises: An utterance from a native speaker, The corresponding image prompt that inspired the utterance and a manually annotated transcription (available for a subset of the data) The dataset has undergone multiple rounds of quality evaluation to ensure reliability and usability for research and deployment. VAANI includes speech data in 86 languages, many of which are low-resource and rarely represented in existing public datasets: Agariya, Angami, Angika, Ao, Assamese, Awadhi, Bagheli, Bagri, Bajjika, Bearybashe, Bengali, Bhatri, Bhili, Bhojpuri, Bihari, Bundeli, Chakhesang, Chakma, Chhattisgarhi, Dorli, Duruwa, English, Galo, Garhwali, Garo, Gondi, Gujarati, Hajong, Halbi, Harauti, Hindi, Jaipuri, Kannada, Khandeshi, Khariboli, Khorth, KhorthKhotta, Khortha, Kokborok, Konkani, Kumaoni, Kurmali, Kurukh, Lambani, Lotha, Magadhi, MagadhiMagahi, Magahi, Maithili, Malayalam, Malvani, Malvi, Marathi, Marwadi, Marwari, Meitei, Mewari, Mewati, Nagamese, Nepali, Nimadi, NissiDafla, Nyishi, Odia, Oriya, Punjabi, Rajasthani, Rajbanshi, Rengma, Sadri, Sangtam, Santali, Shekhawati, Sindhi, Sumi, Surgujia, Surjapuri, Tagin, Tamil, Telugu, Tenyidie, Thethi, Tulu, Urdu, Wagdi, Wancho. This dataset is designed to support a wide range of speech and language technology applications, including: Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) Foundational speech models for Indian languages Speaker identification and verification Language identification systems Speech enhancement and denoising Multimodal Large Language Models (LLMs) Benchmarking and evaluation for Indic language technologies