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 Nepali?
नेपाली in speech, text and translation: figures from the model cards and papers.

Short answer. Partly. Whisper lists Nepali, but its paper counts 0.6 hours of Nepali training audio and reports 47.1% word error rate on read sentences (FLEURS, large-v2). On natural speech a Whisper model scored 100.8% in the IndicVoices paper, against 16.4% for AI4Bharat’s IndicASR. Meta’s MMS lists Nepali for speech recognition but not synthesis. Qwen3 lists Nepali; Meta’s Llama cards do not. IndicTrans2 scored 51.5 chrF++ from English to Nepali on conversational sentences. In the Svarah paper, English spoken by native Nepali speakers was among the hardest for Whisper, at 9.8%. ANVE has not tested Nepali yet. Coverage and figures are as of October 2026.

01

What open models cover in Nepali

TaskModelsWhat the source says
Speech to text: WhisperWhisper large-v2, large-v3The Whisper paper counts 0.6 hours of Nepali speech with Nepali transcripts and 133 hours paired with English text. Large-v2 scored 47.1% word error rate on FLEURS. The large-v3 card gives no Nepali figure.
Speech to text: Indian-language and Meta modelsIndicConformer, Meta MMSIndicConformer lists Nepali (MIT licence); IndicWhisper leaves it out of its 12 languages. IndicVoices, natural speech: IndicASR 16.4% (trained on 100 hours), Whisper 100.8%; the paper gives no MMS figure for Nepali. Meta’s table marks Nepali for speech recognition, not synthesis; the checkpoint is CC-BY-NC.
Chat, summaries, draftingQwen3, Sarvam, Gemma, LlamaQwen3’s launch post lists Nepali. The Llama 3.3 and 4 cards, the gpt-oss card and MILU cover no Nepali, so no chat figure is available. Sarvam 30B and 105B list Nepali (ne) in their card metadata, and AI4Bharat’s IndicBERT v2 includes Nepali.
TranslationIndicTrans2, NLLB-200IndicTrans2 paper (2023), conversational set (IN22-Conv), chrF++: English to Nepali 51.5, Nepali to English 63.0; NLLB 1.2B distilled scored 44.9 and 59.9. On web and Wikipedia sentences (IN22-Gen): 49.0 and 67.7, where Google Translate scored 69.0 into English. NLLB-200 lists Nepali (npi_Deva) and is CC-BY-NC.
02

Devanagari, Nepal and India’s Nepali speakers

Nepali (नेपाली) is written in Devanagari. The 2011 Census counted 29,26,168 people in India with Nepali as their mother tongue, 0.24% of the population and the sixteenth largest group. West Bengal had 11,55,375, Assam 5,96,210, Sikkim 3,82,200, Uttarakhand 1,06,399 and Arunachal Pradesh 95,317. The Census counts Gorkhali under Nepali, and it counts people in India only: Nepali spoken in Nepal is outside these figures. The Home Ministry says Nepali joined the Eighth Schedule in 1992.

Speakers growing up in West Bengal, Sikkim or Assam may mix Nepali with Bengali, Hindi or English, and none of the sources reports a mixed-language Nepali result. The Svarah authors name Nepali, with Assamese and Bodo, among the north-eastern languages whose native speakers Whisper handled worst in English (9.8% against 5.3% for native Hindi speakers).

03

What to test before you rely on it

  • Record unscripted Nepali, not read sentences. Whisper’s 47.1% on read sentences and 100.8% on natural speech differ by more than a factor of two.
  • Compare IndicConformer with Whisper on the same ten minutes; the IndicVoices gap is wide enough to show quickly.
  • Check that the transcript is Nepali and not Hindi; both use Devanagari, so a reader must mark any substitutions.
  • If staff or clients from Nepal are involved, test their speech separately: the Census figures above do not describe it.
  • Translate a notice in both directions and ask a Nepali reader to mark errors; compare IndicTrans2 with NLLB-200 on the same text.
04

Where ANVE stands

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

Because ANVE does not update itself over the internet, the models you test are the models you keep until an update is installed. Today ANVEAI installs updates on site; signed offline USB updates are not on ANVE yet. See the security page.

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

05

Questions people ask

Can Whisper transcribe Nepali?
It accepts Nepali, but with 0.6 hours of Nepali training audio the paper reports 47.1% word error rate on read FLEURS sentences for large-v2, and a Whisper model scored 100.8% on natural speech in the IndicVoices paper.
Is there an offline Nepali speech model?
AI4Bharat’s IndicConformer lists Nepali (MIT licence), and Meta’s MMS lists Nepali speech recognition under a non-commercial licence. In the IndicVoices paper IndicASR scored 16.4% on natural Nepali speech.
Do Llama or Qwen models support Nepali?
Qwen3’s launch post lists Nepali. Meta’s Llama 3.3 and Llama 4 cards list Hindi but not Nepali. No benchmark figure for Nepali chat appears in the cards checked.
Can offline AI translate Nepali?
Yes. IndicTrans2 scored 51.5 chrF++ from English to Nepali and 63.0 from Nepali to English on conversational sentences, against 44.9 and 59.9 for NLLB 1.2B distilled. Have a Nepali 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