ANVE: the private AI office, on the record. See what that means

ANVE

Offline AI in your language · updated 2026-10-07

Can offline AI handle Santali?
ᱥᱟᱱᱛᱟᱲᱤ in Ol Chiki: what open models cover for speech, text and translation.

Short answer. Narrowly. Santali is the one scheduled language in the Austro-Asiatic family, it is written in Ol Chiki script, and it is not in Whisper or Qwen3’s list. Meta’s MMS lists Santali for speech recognition, and AI4Bharat’s IndicConformer lists it; in the IndicVoices paper IndicASR scored 35.4% word error rate on natural Santali speech. IndicTrans2 scored 34.6 chrF++ from English to Santali on conversational sentences, while NLLB 1.2B distilled scored 0.8 in the same direction. Sarvam says its tokenizer is notably more efficient for Santali than other open tokenizers. ANVE has not tested Santali yet. Coverage and figures are as of October 2026.

01

What open models cover in Santali

TaskModelsWhat the source says
Speech to text: WhisperWhisper large-v2, large-v3Santali is not among Whisper’s 100 languages, and the Whisper paper reports no Santali result.
Speech to text: Meta and Indian-language modelsMeta MMS, AI4Bharat IndicConformerMeta’s MMS table marks Santali (“Santhali”) for speech recognition, not synthesis; the checkpoint is CC-BY-NC. IndicConformer lists Santali (sat; MIT licence). IndicVoices, natural speech: IndicASR 35.4%, trained on 86 hours; the paper gives no Whisper or MMS figure for it. IndicWhisper leaves Santali out of its 12 languages.
Chat, summaries, draftingSarvam, IndicBERT-v3, Qwen3, LlamaQwen3’s list has no Santali, and the Llama 3.3 and 4 cards, the gpt-oss card and MILU cover none. Sarvam 30B and 105B list Santali (sat) in their card metadata, and Sarvam’s post says its tokenizer is significantly more efficient for Santali than other tokenizers. AI4Bharat’s IndicBERT-v3 adds Santali in a low-resource phase (an encoder, not a chat model).
TranslationIndicTrans2, NLLB-200, Sarvam-TranslateIndicTrans2 paper (2023), Ol Chiki, chrF++: conversational set (IN22-Conv) English to Santali 34.6, Santali to English 43.5; NLLB 1.2B distilled 0.8 and 33.6; NLLB 54B mixture-of-experts 18.0 and 37.4. On web and Wikipedia sentences (IN22-Gen): IndicTrans2 33.4 and 45.3, with NLLB 1.2B distilled’s English-to-Santali output marked completely off-target (about 1.0). On FLORES-200 IndicTrans2 scored 28.4 and 39.3, and NLLB 54B 27.0 and 44.3, ahead for Santali to English. Google Translate and Azure have no Santali result in the 2023 IndicTrans2 tables. NLLB-200 lists Santali (sat_Olck) and is CC-BY-NC.
02

Ol Chiki, three states and an Austro-Asiatic language

Santali (ᱥᱟᱱᱛᱟᱲᱤ) is written in Ol Chiki in IndicTrans2 and FLORES-200 (sat_Olck). The IndicTrans2 README counts it among the five scripts the model supports, alongside Devanagari, Perso-Arabic, Meitei and Latin. The 2011 Census counted 73,68,192 people with Santali as their mother tongue, 0.61% of India and the fourteenth largest group, up 13.89% from 2001. Jharkhand had 32,69,897, West Bengal 24,29,073, Odisha 8,62,590, Bihar 4,58,949, Assam 2,13,139 and Maharashtra 1,03,456. The Census groups Karmali (3,58,579) and Mahili (26,399) under Santali, and counts Santali as the only scheduled language in the Austro-Asiatic family, the Indo-Aryan, Dravidian and Tibeto-Burmese languages being the rest. The Home Ministry spells it Santhali and says it joined the Eighth Schedule in 2004.

Script is a practical question. The open models above list Ol Chiki, so check whether your office’s fonts, keyboards and document formats handle Ol Chiki before you test accuracy; none of the sources tests document tools.

The Svarah paper’s 19 mother-tongue groups do not include Santali, so there is no accent result for Santali speakers’ English.

03

What to test before you rely on it

  • Decide the script, and check that your document tools can display and print it.
  • For speech, compare IndicConformer with Meta’s MMS (if its non-commercial licence suits you) on ten minutes of your own natural speech.
  • For translation, compare IndicTrans2 and NLLB-200 on the same text in both directions: in the paper NLLB 54B was ahead for Santali to English on FLORES-200, and the smaller NLLB 1.2B model failed English to Santali.
  • Check names of villages, clans and festivals, and numbers and dates said aloud.
  • Have a Santali reader mark errors in any translated notice; automatic scores in the paper are overlap measures, not human ratings.
04

Where ANVE stands

ANVE’s meeting-minutes maker handles Hindi, English and Hinglish today. Santali has not been tested on ANVE yet, so there is no ANVE result to report and nothing here promises support.

ANVE can call a cloud model only if an administrator adds one, and the answers are labelled as cloud answers. For a language like Santali, where the local evidence above is thin, that is a decision for the office to take deliberately, with every call logged.

A demo call is free. If Santali decides your purchase, say which script you need and ask for your own recordings to be tried before you commit.

05

Questions people ask

Does Whisper support Santali?
No. Santali is not in Whisper’s list of 100 languages and the paper reports no Santali result. Meta’s MMS and AI4Bharat’s IndicConformer list it.
Which script does offline Santali AI use?
Ol Chiki. AI4Bharat’s IndicTrans2 and FLORES-200 list Santali as sat_Olck. Check that your document tools support it before you test accuracy.
Is there an offline Santali speech model?
AI4Bharat’s IndicConformer lists Santali (MIT licence), and Meta’s MMS lists Santali speech recognition under a non-commercial licence. In the IndicVoices paper IndicASR scored 35.4% on natural Santali speech.
Can offline AI translate Santali?
Yes, with lower scores than for larger languages. IndicTrans2 scored 34.6 chrF++ from English to Santali and 43.5 from Santali to English on conversational sentences. NLLB 1.2B distilled scored 0.8 from English, so compare the models on your text.
06

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

07

Sources