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

ANVE

Offline AI in your language · updated 2026-10-16

Can offline AI handle Maithili?
मैथिली in speech, text and translation: thin coverage, but not none.

Short answer. Possible, but thin. Whisper does not include Maithili. Meta’s MMS lists Maithili for speech recognition, under a non-commercial licence, and AI4Bharat’s IndicConformer lists it; in the IndicVoices paper AI4Bharat’s IndicASR scored 33.6% word error rate on natural Maithili speech against 77.1% for MMS. Qwen3 lists Maithili for text, and Meta’s Llama cards do not. IndicTrans2 scored 44.3 chrF++ from English to Maithili on conversational sentences. The Census counted Maithili under Hindi until 1991, and the Home Ministry added it to the Eighth Schedule in 2004. ANVE has not tested Maithili yet. Coverage and figures are as of October 2026.

01

What open models cover in Maithili

TaskModelsWhat the source says
Speech to text: WhisperWhisper large-v2, large-v3Maithili is not among the 100 languages in Whisper’s tokenizer, and the Whisper paper reports no Maithili result.
Speech to text: Indian-language and Meta modelsIndicConformer, Meta MMSIndicConformer lists Maithili (MIT licence). Meta’s MMS table marks Maithili for speech recognition and synthesis; the checkpoint is CC-BY-NC. IndicWhisper leaves Maithili out of its 12 languages. IndicVoices, natural speech: IndicASR 33.6% (trained on 100 hours), MMS 77.1%.
Chat, summaries, draftingQwen3, Sarvam, AI4Bharat IndicBERT-v3Qwen3’s launch post lists Maithili. The Llama 3.3 and 4 cards, the gpt-oss card and MILU cover no Maithili, so no chat figure is available. Sarvam 30B and 105B list Maithili (mai) in their card metadata. AI4Bharat’s IndicBERT-v3 adds Maithili in a low-resource training phase; it is an encoder for classification and search, not a chat model.
TranslationIndicTrans2, NLLB-200IndicTrans2 paper (2023), conversational set (IN22-Conv), chrF++: English to Maithili 44.3, Maithili to English 57.8; NLLB 1.2B distilled scored 41.6 and 55.2. On web and Wikipedia sentences (IN22-Gen): 48.7 and 64.8; on FLORES-200: 50.5 and 69.5. NLLB-200 lists Maithili (mai_Deva) and is CC-BY-NC.
02

Devanagari, Bihar and a language once counted as Hindi

Maithili (मैथिली) is written in Devanagari in all the sources above. The 2011 Census counted 1,35,83,464 people with Maithili as their mother tongue, 1.12% of India and the thirteenth largest group. Bihar had 1,30,63,042 of them. Jharkhand had 1,38,167, Delhi 1,22,956 and Uttarakhand 54,553. The Census groups Thati (1,65,420 speakers) and Tharu (53,575) with Maithili.

Until the 1991 Census, Maithili speakers were counted as speakers of Hindi, and the Home Ministry says Maithili joined the Eighth Schedule in 2004. For a speech or text model, Maithili and Hindi are separate entries, but a Hindi model may still hear or write Hindi where Maithili was spoken. None of the sources tests that.

For English meetings, the Svarah paper found Whisper large at 4.5% word error rate on speakers whose mother tongue is Maithili, the lowest of the 19 mother-tongue groups it reports (native Hindi speakers: 5.3%).

03

What to test before you rely on it

  • Check that the transcript is Maithili and not Hindi: ask a Maithili reader to mark words that the model has “corrected” into Hindi.
  • Try IndicConformer first. Meta’s MMS is CC-BY-NC, which bars commercial use, and Whisper cannot be told to transcribe Maithili.
  • Record speakers from Bihar and from Jharkhand or Delhi, and include phone-quality audio if your callers use it.
  • Check names, rupee amounts and dates said aloud, and that Devanagari spellings match your documents.
  • For translation, run the same notice through IndicTrans2 and NLLB-200 and ask a Maithili reader which keeps the meaning.
04

Where ANVE stands

ANVE is pre-release, and its meeting-minutes maker handles Hindi, English and Hinglish in the development build. Maithili is not among them and has not been tested, so treat the table above as the whole evidence.

A cloud model is optional on ANVE: it runs only after an administrator adds a provider, its answers are labelled, and each call is logged as local or cloud. A policy that blocks confidential data from cloud models is not on ANVE yet, so until it ships treat any cloud route for Maithili as a conscious exception.

A demo call is free. If Maithili decides your purchase, ask for your own recordings to be tried before you commit.

05

Questions people ask

Does Whisper support Maithili?
No. Maithili is not in Whisper’s list of 100 languages, and the paper reports no Maithili result.
Which offline model can transcribe Maithili?
AI4Bharat’s IndicConformer lists Maithili under an MIT licence, and Meta’s MMS lists it under a non-commercial licence. In the IndicVoices paper IndicASR scored 33.6% on natural Maithili speech and MMS 77.1%.
Is Maithili the same as Hindi?
The Census has counted Maithili separately since 2001 and the Home Ministry lists it as its own Eighth Schedule language, added in 2004. Until 1991 the Census counted it within Hindi. Models treat them as separate languages.
Can offline AI translate Maithili?
Yes. IndicTrans2 scored 44.3 chrF++ from English to Maithili and 57.8 from Maithili to English on conversational sentences, against 41.6 and 55.2 for NLLB 1.2B distilled. Have a Maithili reader check anything official.
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