vnahata/vaani-audio-image-retrieval

Vaani audio–image retrieval (MTEB)

Multilingual audio↔image retrieval over 62 Indian languages, derived from Project Vaani (IISc Bangalore / ARTPARK).

Vaani records image-prompted speech: a speaker is shown a photograph and describes it aloud in their own language. Each recording is therefore grounded in a specific image, which is what makes audio↔image retrieval well defined without any extra annotation.

Prepared for MTEB as VaaniA2IRetrieval and VaaniI2ARetrieval.

Contents

config	rows	description
images	3,552	shared image corpus
one per language (62)	25–150 each	recordings, 6,056 total

Language configs carry id, audio, image_id, language, speaker_id, gender, state, district, duration, transcript. image_id joins to the images config.

Languages span Indo-Aryan, Dravidian, Tibeto-Burman and Munda — Chhattisgarhi, Malayalam, Urdu, Kashmiri, Karbi, Wancho, Kokborok, Santali, Tulu, Lepcha, Sikkimese and others.

How this was built

The source is ~3.8 TB of per-language audio shards plus a 49 GB pool of 289,676 images, in separate trees, linked only by a referenceImage filename on each audio row. The build indexes the image pool by filename — reading only the path column, never the image bytes — then fetches just the images the sampled audio references.

Sampling is speaker-capped and spread across row groups. Vaani records long sessions per speaker, so a contiguous read returns the same voice repeatedly. Clips outside a 1.5–30s duration band are dropped, as are languages left with fewer than 25 usable clips (62 of 84 kept). Byte-identical duplicate recordings and images are removed.

This is an evaluation sample, not a copy of the corpus. The source is 31,255 hours; this subset is sized for audio encoding cost.

Split naming: Vaani ships only a train split. These are evaluation-only tasks, so the sampled evaluation set is published here as test. No training data is retained.

Construction script: scripts/data/vaani_retrieval/create_data.py in the MTEB repo.

Known limitations

• A handful of Vaani's language labels (Bajjika, Bearybashe, Khortha, Malvani, Pahadi) have no distinct ISO 639-3 entry and are mapped to the closest coded variety; these are marked in the task metadata.
• Transcripts are present for only ~10% of clips and are not used by these tasks.

Citation

@misc{pulikodan2026vaani,
  archiveprefix = {arXiv},
  author = {Pulikodan, Sujith and Singh, Abhayjeet and Basu, Agneedh and Desai, Nihar and J, Pavan Kumar and Bhat, Pranav D and Dharmaraju, Raghu and Gupta, Ritika and Udupa, Sathvik and Kumar, Saurabh and Sharma, Sumit and Sanka, Visruth and Tewari, Dinesh and Dhand, Harsh and Kamat, Amrita and Singh, Sukhwinder and Vashishth, Shikhar and Talukdar, Partha and Acharya, Raj and Ghosh, Prasanta Kumar},
  eprint = {2603.28714},
  title = {VAANI: Capturing the language landscape for an inclusive digital India},
  year = {2026},
}

Redistributed under CC-BY-4.0, the licence of the source dataset.