Can Whisper transcribe Sanskrit?

It has a Sanskrit tag, but the paper lists no Sanskrit transcription data, only 195 hours paired with English text, and a Whisper model scored 99.0% word error rate on natural Sanskrit in the IndicVoices paper.

Is there an offline Sanskrit speech model?

Yes. AI4Bharat’s IndicWhisper has a Sanskrit model (MIT licence), averaging 48.0% on its test sets, and IndicConformer lists Sanskrit. In the IndicVoices paper IndicASR scored 22.8% on natural Sanskrit speech.

Can offline AI translate Sanskrit?

Yes, with modest scores. IndicTrans2 scored 35.5 chrF++ from English to Sanskrit and 48.3 from Sanskrit to English on conversational sentences, ahead of NLLB 1.2B distilled at 26.4 and 38.9. Have a Sanskrit scholar check anything that matters.

Does Sanskrit appear in the Census?

Yes: 24,821 people gave it as their mother tongue in 2011, which the Census marks as negligible as a share of the population. It is also one of the 22 languages in the Constitution’s Eighth Schedule.

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 Sanskrit?
संस्कृतम् in speech, text and translation: a classical language with few speakers and thin data.

Short answer. Thinly. Only 24,821 people gave Sanskrit as their mother tongue in the 2011 Census. Whisper lists Sanskrit, but its paper has no Sanskrit transcription data, only 195 hours of Sanskrit speech paired with English text, and a Whisper model scored 99.0% word error rate on natural Sanskrit in the IndicVoices paper against 22.8% for AI4Bharat’s IndicASR. IndicWhisper’s Sanskrit average, 48.0%, is the highest of its 12 languages. IndicTrans2 scored 35.5 chrF++ from English to Sanskrit on conversational sentences, well above NLLB 1.2B distilled at 26.4. Qwen3 does not list Sanskrit. ANVE has not tested Sanskrit yet. Coverage and figures are as of October 2026.

What open models cover in Sanskrit

Task	Models	What the source says
Speech to text: Whisper	Whisper large-v2, large-v3	Sanskrit (sa) is in Whisper’s tokenizer, but the paper lists no Sanskrit speech with Sanskrit transcripts: its only Sanskrit data is 195 hours of Sanskrit speech paired with English text. Sanskrit is not in FLEURS, so the paper reports no Sanskrit word error rate.
Speech to text: Indian-language models	IndicWhisper, IndicConformer	IndicWhisper’s Sanskrit model (MIT licence), trained on 207 hours of Sanskrit, averages 48.0% across its Sanskrit test sets, the highest of its 12 languages. IndicConformer lists Sanskrit (sa). IndicVoices, natural speech: IndicASR 22.8% (trained on 45 hours), Whisper 99.0%; Meta’s MMS lists Sanskrit for language identification only.
Chat, summaries, drafting	Sarvam, IndicTrans3-beta, IndicBERT-v3, Qwen3	Qwen3’s list has no Sanskrit, and the Llama 3.3 and 4 cards, the gpt-oss card and MILU cover none. Sarvam 30B and 105B list Sanskrit (sa) in their card metadata. IndicTrans3-beta puts Sanskrit in its primary set of 15 languages, and AI4Bharat’s IndicBERT-v3 adds it in a low-resource phase (an encoder, not a chat model).
Translation	IndicTrans2, NLLB-200, Sarvam-Translate	IndicTrans2 paper (2023), conversational set (IN22-Conv), chrF++: English to Sanskrit 35.5, Sanskrit to English 48.3; NLLB 1.2B distilled scored 26.4 and 38.9; Google Translate 32.8 and 49.2. On web and Wikipedia sentences (IN22-Gen): 38.8 and 54.8. NLLB-200 lists Sanskrit (san_Deva) and is CC-BY-NC. Sarvam-Translate lists Sanskrit.

Devanagari, a tiny Census count and classical text

Sanskrit (संस्कृतम्) is written in Devanagari in the sources above (san_Deva). The 2011 Census counted 24,821 people with Sanskrit as their mother tongue, a share it marks as negligible, and ranks it last of the 22 scheduled languages. The counts swing: 49,736 in 1991, 14,135 in 2001 and 24,821 in 2011. Maharashtra had 3,802, Bihar 3,388, Uttar Pradesh 3,062, Rajasthan 2,375 and Madhya Pradesh 1,871. Sanskrit is one of the languages the Constitution listed from the start.

Most Sanskrit text is classical or ritual: verse, commentary and chant. The Sanskrit speech data in AI4Bharat’s Vistaar collection includes read speech of pre-classical, classical and modern Sanskrit literature, and 207 hours in all. A speech model trained on that may not follow a recited verse, a lecture mixed with Hindi or a Sanskrit sentence spoken at conversation speed; none of the sources tests those.

There is no Svarah accent result for Sanskrit. For translation, treat the paper’s scores as direction: 35.5 chrF++ into Sanskrit is well below the 49.6 the same paper reports into Hindi.

What to test before you rely on it

State what you need: transcripts of spoken Sanskrit, translation of Sanskrit texts into English or Hindi, or answers about them. The evidence differs for each.
For speech, start with IndicWhisper’s Sanskrit model or IndicConformer, and record the kind of speech you have: recited verse, lecture or conversation.
Check how the output handles joined words and sandhi, and whether names and citations keep the spellings your readers expect.
Translate a short passage in both directions and ask a Sanskrit scholar to mark errors. The translation scores are overlap measures, not human ratings.
Treat chat answers about Sanskrit texts as unverified: no benchmark figure for Sanskrit chat appears in the cards checked.

Where ANVE stands

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

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 Sanskrit as a conscious exception.

A demo call is free. If Sanskrit decides your purchase, bring a sample of the speech or text you have and ask for it to be tried before you commit.

Questions people ask

Can Whisper transcribe Sanskrit?: It has a Sanskrit tag, but the paper lists no Sanskrit transcription data, only 195 hours paired with English text, and a Whisper model scored 99.0% word error rate on natural Sanskrit in the IndicVoices paper.
Is there an offline Sanskrit speech model?: Yes. AI4Bharat’s IndicWhisper has a Sanskrit model (MIT licence), averaging 48.0% on its test sets, and IndicConformer lists Sanskrit. In the IndicVoices paper IndicASR scored 22.8% on natural Sanskrit speech.
Can offline AI translate Sanskrit?: Yes, with modest scores. IndicTrans2 scored 35.5 chrF++ from English to Sanskrit and 48.3 from Sanskrit to English on conversational sentences, ahead of NLLB 1.2B distilled at 26.4 and 38.9. Have a Sanskrit scholar check anything that matters.
Does Sanskrit appear in the Census?: Yes: 24,821 people gave it as their mother tongue in 2011, which the Census marks as negligible as a share of the population. It is also one of the 22 languages in the Constitution’s Eighth Schedule.

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.

Can offline AI handle Sanskrit?संस्कृतम् in speech, text and translation: a classical language with few speakers and thin data.