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 Assamese?
অসমীয়া in speech, text and translation: what the sources report and what is missing.

Short answer. Partly: translation is workable, speech needs an Indian-language model, and chat is thinly evidenced. Whisper has an Assamese tag, but its paper lists no Assamese speech with Assamese transcripts, only 136 hours of Assamese speech paired with English text, and it scored 106.2% word error rate on FLEURS (large-v2). In the IndicVoices paper AI4Bharat’s IndicASR scored 20.4% on natural Assamese speech, Meta’s MMS 51.0% and a Whisper model 127.2%. Qwen3 lists Assamese; Meta’s Llama cards do not. IndicTrans2 scored 46.8 chrF++ from English to Assamese on conversational sentences. ANVE has not tested Assamese yet. Coverage and figures are as of October 2026.

01

What open models cover in Assamese

TaskModelsWhat the source says
Speech to text: WhisperWhisper large-v2, large-v3Whisper’s tokenizer includes Assamese (as), but the paper lists no Assamese speech with Assamese transcripts. Its only Assamese data is 136 hours of Assamese speech paired with English text. Large-v2 scored 106.2% word error rate on FLEURS.
Speech to text: Indian-language modelsIndicConformer, Meta MMSIndicConformer lists Assamese (MIT licence); IndicWhisper leaves it out of its 12 languages. IndicVoices, natural speech: IndicASR 20.4% (trained on 100 hours), MMS 51.0%, Whisper 127.2%. Meta’s table marks Assamese for speech recognition and synthesis; the checkpoint is CC-BY-NC.
Chat, summaries, draftingQwen3, Sarvam, Gemma, LlamaQwen3’s launch post lists Assamese. The Llama 3.3 and 4 cards, the gpt-oss card and MILU cover no Assamese, so no chat figure is available. Sarvam 30B and 105B list Assamese (as) in their card metadata, but Sarvam’s post says most multilingual training went to the 10 most-spoken Indian languages and does not name them; Assamese ranked 12th in the 2011 Census.
TranslationIndicTrans2, NLLB-200IndicTrans2 paper (2023), conversational set (IN22-Conv), chrF++: English to Assamese 46.8, Assamese to English 62.9; NLLB 1.2B distilled scored 42.6 and 58.7. On web and Wikipedia sentences (IN22-Gen): 47.1 and 65.8. NLLB-200 lists Assamese (asm_Beng) and is CC-BY-NC.
02

Assamese script, Assam and the north-east

Assamese (অসমীয়া) is written in a script that FLORES-200 and IndicTrans2 label with the same code as Bengali (asm_Beng). It has letters Bengali does not use, such as ৰ and ৱ. The 2011 Census counted 1,53,11,351 people with Assamese as their mother tongue, 1.26% of India and the twelfth largest group, up 16.27% in the decade. Assam had 1,50,95,797; Arunachal Pradesh had 53,951 and Meghalaya 39,628.

In Assam the same Census table also counted 90,24,324 Bengali and 14,16,125 Bodo speakers, so one office can need all three. No source tests mixed Assamese, Bengali and Bodo speech.

For English meetings, the Svarah paper found Whisper large at 10.1% word error rate on speakers whose mother tongue is Assamese, against 5.3% for native Hindi speakers. Its authors say models do worse for speakers of Assamese, Bodo and Nepali, low-resource languages of the north-east.

03

What to test before you rely on it

  • Test Assamese on its own, not as a form of Bengali. Check that the output uses Assamese spellings, including ৰ and ৱ, and that digits come out as you expect.
  • Use an Indian-language speech model such as IndicConformer as the starting point. Whisper’s 106.2% on read sentences and 127.2% on natural speech make it a poor first try.
  • Record Assamese and Bengali speakers in the same meeting and check the model keeps them apart; AI4Bharat’s IndicLID identifies both in native and romanised text.
  • Run your own documents through a chat model and ask an Assamese reader to mark errors. No source gives a benchmark figure to lean on.
  • Translate a notice in both directions and compare IndicTrans2 with NLLB-200 on the same text.
04

Where ANVE stands

On ANVE, only Hindi, English and Hinglish are handled by the meeting-minutes maker today. Assamese is untested there: no result, and no promise of 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 Assamese, 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 Assamese decides your purchase, ask for your own recordings to be tried before you commit.

05

Questions people ask

Does Whisper support Assamese?
It has an Assamese language tag, but the paper lists no Assamese speech with Assamese transcripts, only 136 hours paired with English text. Large-v2 scored 106.2% word error rate on FLEURS.
Is there an offline Assamese speech model?
AI4Bharat’s IndicConformer lists Assamese (MIT licence), and Meta’s MMS does too under a non-commercial licence. In the IndicVoices paper IndicASR scored 20.4% on natural Assamese speech.
Is Assamese the same as Bengali for AI models?
No. They share a script code in FLORES-200 and IndicTrans2, but every list treats them as separate languages with separate test sets, and Assamese uses letters Bengali does not.
Can offline AI translate Assamese?
Yes. IndicTrans2 scored 46.8 chrF++ from English to Assamese and 62.9 from Assamese to English on conversational sentences, against 42.6 and 58.7 for NLLB 1.2B distilled. Have an Assamese 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