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 Kashmiri?
Koshur in speech, text and translation: AI4Bharat’s models in two scripts, little else.

Short answer. Barely. Kashmiri is not in Whisper, not in Qwen3’s language list and not in Meta’s MMS for speech recognition. AI4Bharat’s IndicConformer lists it, and IndicASR scored 39.3% word error rate on natural Kashmiri speech in the IndicVoices paper, the second-highest of its 22 languages, after 50 hours of training. For translation IndicTrans2 and NLLB-200 both cover Kashmiri in Arabic and Devanagari scripts, but the script matters: on FLORES-200 IndicTrans2 scored 39.7 chrF++ from English into Arabic-script Kashmiri and only 19.2 into Devanagari-script Kashmiri. ANVE has not tested Kashmiri yet. Coverage and figures are as of October 2026.

01

What open models cover in Kashmiri

TaskModelsWhat the source says
Speech to text: Whisper, MMSWhisper, Meta MMSKashmiri is not among Whisper’s 100 languages, and the paper reports no Kashmiri result. Meta’s MMS table marks Kashmiri for language identification only, with no speech recognition or synthesis.
Speech to text: Indian-language modelsAI4Bharat IndicConformerIndicConformer lists Kashmiri (ks; MIT licence). IndicVoices, natural speech: IndicASR 39.3%, trained on 50 hours, the second-highest error rate of its 22 languages; the paper gives no Whisper or MMS figure. IndicWhisper leaves Kashmiri out of its 12 languages.
Chat, summaries, draftingSarvam, IndicBERT-v3, Qwen3, LlamaQwen3’s list has no Kashmiri, and the Llama 3.3 and 4 cards, the gpt-oss card and MILU cover none. Sarvam 30B and 105B list Kashmiri (ks) in their card metadata; AI4Bharat’s IndicBERT-v3 adds it in a low-resource training phase and IndicTrans3-beta gives it preliminary support. IndicBERT-v3 is an encoder for classification and search, not a chat model.
TranslationIndicTrans2, NLLB-200FLORES-200, chrF++, English to Kashmiri: Arabic script 39.7 for IndicTrans2 and 37.2 for NLLB 1.2B distilled; Devanagari script 19.2 and 18.7. Kashmiri to English: 59.7 and 57.8 in Arabic script, 48.3 and 47.7 in Devanagari. On conversational sentences (IN22-Conv, Arabic script): 35.6 and 52.6 against 25.7 and 44.6. Google Translate and Azure have no Kashmiri results in the 2023 IndicTrans2 tables.
02

Two scripts and the Valley

Kashmiri (Koshur) is written in Perso-Arabic script and in Devanagari; IndicTrans2 and NLLB-200 list both (kas_Arab and kas_Deva). The 2011 Census counted 67,97,587 people with Kashmiri as their mother tongue, 0.56% of India and the fifteenth largest group, up 22.97% from 2001. Jammu and Kashmir had 66,80,837 of them. Himachal Pradesh had 57,050 and Delhi 18,122. The Census groups Siraji (1,24,896), Kishtwari (39,748) and Dardi (25,600) under Kashmiri. By the Home Ministry’s account of later additions, Kashmiri is one of the 14 languages the Constitution listed from the start.

The script gap is the warning. On the same benchmark IndicTrans2 scored 39.7 into Arabic-script Kashmiri and 19.2 into Devanagari, so a model that is competent in one script may be close to unusable in the other. Decide which script your office writes before you test anything else.

For English meetings, the Svarah paper found Whisper large at 6.9% word error rate on speakers whose mother tongue is Kashmiri, against 5.3% for native Hindi speakers.

03

What to test before you rely on it

  • Choose the script, then test only that script. Ask a Kashmiri reader to judge a translated notice in both directions.
  • For speech, start with IndicConformer on ten minutes of your own recordings. IndicASR’s 39.3% was measured on natural speech, so expect a lot of corrections.
  • Check names of places and families, which the training data are least likely to contain, and numbers and dates said aloud.
  • Test speakers of the varieties the Census groups under Kashmiri, such as Kishtwari and Siraji, if your clients use them; none of the sources reports results for them.
  • Treat any chat answer in Kashmiri as unverified until a reader has checked it: no benchmark figure appears in the cards checked.
04

Where ANVE stands

On ANVE, only Hindi, English and Hinglish are handled by the meeting-minutes maker today. Kashmiri 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 Kashmiri, 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 Kashmiri decides your purchase, ask for your own recordings to be tried before you commit.

05

Questions people ask

Does Whisper support Kashmiri?
No. Kashmiri is not in Whisper’s list of 100 languages and the paper reports no Kashmiri result. AI4Bharat’s IndicConformer lists Kashmiri instead.
Which script do Kashmiri models use?
IndicTrans2 and NLLB-200 list both Arabic script and Devanagari, but on FLORES-200 IndicTrans2 scored 39.7 into Arabic script and only 19.2 into Devanagari. The speech model card does not state a script.
Is there an offline Kashmiri speech model?
AI4Bharat’s IndicConformer lists Kashmiri (MIT licence). In the IndicVoices paper IndicASR scored 39.3% on natural Kashmiri speech, trained on 50 hours.
Can offline AI translate Kashmiri?
Yes, in Arabic script the scores are moderate: 39.7 chrF++ from English and 59.7 into English on FLORES-200. Devanagari-script Kashmiri scored far lower. Have a Kashmiri 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