Tulu AI Training Datasets
Tulu (ತುಳು) spoken in Karnataka / Kerala (South India), written in Kannada / Tigalari, is one of the languages we build proprietary, rights-cleared datasets for. It is a low-resource language: usable training data barely exists on the open web, so it has to be collected from real speakers with consent. Every Tulu dataset ships with documented consent, provenance and published quality metrics, and is available with India data residency.
Tulu datasets in the catalog
Showing 1 of 41 datasets.
-
GeneralGeneral Audio / Speech 🇮🇳 Sovereign
Rare-Language Conversational Speech Corpus
Spontaneous, real-world conversational speech in four rare Indian languages — transcribed, segmented, and ready for ASR and speech-LLM training.
- Languages:
- Santali, Gondi, Tulu, Bhojpuri
- Size:
- 260 hrs · 1,850 speakers
- IAA:
- 0.85
Rights-clearedLicence from
Contact us
Services for Tulu data
When the data you need doesn't exist yet, we build it — collection, annotation, alignment and evaluation, all rights-cleared and documented.
Multilingual Data Collection
Native-speaker audio, video, image and text collection across Indic, African and low-resource languages — field-grade, consented, documented.
Explore the serviceData Annotation
Speech, NLP, CV and multimodal annotation at IAA ≥ 0.85 with two-pass QA — built for foundation-model SLAs.
Explore the serviceSFT Gold-Standard Data
High-quality prompt/response curation for supervised fine-tuning — vetted by linguistic SMEs, not crowd-sourced noise.
Explore the serviceCultural & Cross-Lingual Evaluation
Evaluation for honorifics, code-mix, idioms, caste-safety and pragmatic correctness — beyond translated MMLU.
Explore the serviceCommon questions
- Is the Tulu data rights-cleared and safe to train on?
Yes. Every Tulu record is created or sourced under written contributor agreements granting commercial reuse, with a consent reference and authorship log — never scraped.
- Can you collect more Tulu data to spec?
Yes. When the Tulu data you need doesn't exist, we collect and annotate it to your specification across speech, text and multimodal modalities, with an India-residency option.
- How is Tulu data quality measured?
Each dataset ships with a data card: inter-annotator agreement, QA pass rate, and (for speech) word error rate, with two-pass QA and adjudication.
Need Tulu data your model can train on?
License what's in the catalog, or tell us exactly what you need — we build proprietary, rights-cleared datasets to spec, with India data residency.