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 Marathi?
मराठी for speech, text and translation: what the sources report and what to check first.

Short answer. Text and translation, yes; speech with the right model, not stock Whisper. Whisper lists Marathi, but its paper counts only 0.6 hours of Marathi training audio and reports 38.3% word error rate on read sentences (FLEURS, large-v2). On natural speech in the IndicVoices paper a Whisper model scored 95.2%, while AI4Bharat’s IndicASR scored 18.2%. Qwen3 lists Marathi; Meta’s Llama cards do not. IndicTrans2 scored 48.6 chrF++ from English to Marathi on conversational sentences. ANVE has not tested Marathi yet. Coverage and figures are as of October 2026.

01

What open models cover in Marathi

TaskModelsWhat the source says
Speech to text: WhisperWhisper large-v2, large-v3The Whisper paper counts 0.6 hours of Marathi speech with Marathi transcripts and 288 hours paired with English text. Large-v2 scored 38.3% word error rate on FLEURS. That is far lower than the paper’s Bengali result (104.1%), although Marathi had less data; the paper does not explain the difference.
Speech to text: Indian-language modelsIndicConformer, IndicWhisper, Meta MMSIndicConformer lists Marathi (MIT licence). IndicWhisper averages 18.2% on Marathi test sets, 20.5% on FLEURS. IndicVoices, natural speech: IndicASR 18.2%, MMS 52.7%, Whisper 95.2%. MMS is CC-BY-NC.
Chat, summaries, draftingQwen3, Gemma, Sarvam, LlamaQwen3’s launch post lists Marathi. The Llama 3.3 and 4 cards do not, and the gpt-oss card tests no Marathi. MILU (2024 test, Indian exam questions; open base models, 5-shot): Llama 3.1 70B 62.2%, Gemma 2 27B 59.66%. Sarvam-1 and Sarvam-M name Marathi among their 10 Indian languages.
TranslationIndicTrans2, NLLB-200IndicTrans2 paper (2023), conversational set (IN22-Conv), chrF++: English to Marathi 48.6, Marathi to English 58.5; NLLB 1.2B distilled 44.8 and 56.2. On IN22-Gen: 51.0 and 63.7.
02

Devanagari, Maharashtra and the neighbours

Marathi (मराठी) is written in Devanagari, the same script as Hindi. The 2011 Census counted 8,30,26,680 people with Marathi as their mother tongue, 6.86% of India and the third largest group. Maharashtra had 7,74,61,172. Karnataka had 20,64,906, Madhya Pradesh 12,31,285, Gujarat 9,20,345 and Goa 1,58,787.

A shared script does not make the languages interchangeable. Every list above treats Hindi and Marathi as separate languages with separate test sets, and a Hindi model is not a substitute for a Marathi one. If your office mixes Hindi, Marathi and English in one meeting, test the mix; none of the sources reports a result for it.

For English meetings, the Svarah paper measured Whisper large at 6.8% word error rate on speakers whose mother tongue is Marathi.

03

What to test before you rely on it

  • Record spontaneous Marathi speech, not read sentences. The paper’s 38.3% is for read sentences; natural speech is a different test.
  • Check Marathi digits and amounts (१ २ ३ against 1 2 3), dates, and the names of places and people.
  • Check that the output stays in Marathi and does not drift into Hindi vocabulary or Devanagari spellings that belong to Hindi.
  • Speak to the model in the way your staff really talk: mixed Marathi and English sentences, or Marathi with Hindi words.
  • Translate a notice in both directions and ask a Marathi reader to mark errors. NLLB-200’s card says its translations cannot be used as certified translations.
04

Where ANVE stands

ANVE’s meeting-minutes maker handles Hindi, English and Hinglish today. Marathi 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 Marathi decides your purchase, ask for your own recordings to be tried on the device before you commit, and judge the transcript line by line.

05

Questions people ask

Can Whisper transcribe Marathi?
It accepts Marathi, and large-v2 scored 38.3% word error rate on read FLEURS sentences despite only 0.6 hours of Marathi training audio. On natural speech the IndicVoices paper put a Whisper model at 95.2%.
Which offline model is most accurate for Marathi speech?
No source ranks every model. In the IndicVoices paper AI4Bharat’s IndicASR scored 18.2% on natural Marathi speech against 52.7% for MMS and 95.2% for Whisper. AI4Bharat’s released IndicConformer models cover Marathi, but their card gives a figure only for Hindi.
Do Llama models support Marathi?
Not officially: Meta’s Llama 3.3 and Llama 4 cards list Hindi but not Marathi. Qwen3 lists Marathi, and Sarvam’s smaller models name it among 10 Indian languages.
Can offline AI translate between Marathi and English?
Yes. IndicTrans2 scored 48.6 chrF++ English to Marathi and 58.5 Marathi to English on conversational sentences, higher than the NLLB 1.2B distilled model at 44.8 and 56.2. Check anything official with a human reader.
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