How accurate is offline speech recognition on Indian-accented English?

On the Svarah benchmark Whisper large scored 7.2% word error rate against 1.6% on a native-speaker read set, and 11.2% on everyday prompts full of brand names and IDs. The paper dates from 2023 and tested the original Whisper base, medium and large models.

Does the speaker’s mother tongue change accuracy?

Yes. In the Svarah paper Whisper large scored 4.5% on Maithili speakers’ English and 5.3% on Hindi speakers’, but 9.8% on Nepali, 10.1% on Assamese and 11.6% on Bodo speakers’. Test your own speakers.

Should I use a bigger Whisper model for Indian accents?

Bigger did better in the paper: base 13.6%, medium 8.3%, large 7.2%. Whisper’s README lists about 1 GB, 5 GB and 10 GB of video memory for those sizes, so match the model to the hardware you have.

Is English counted in the Indian Census?

Yes, as a mother tongue: 2,59,678 people in 2011, or 0.02%. Most English use in India is as a second or later language, which the mother-tongue table does not capture.

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 Indian English?
Accents, names, rupee amounts and Hindi words in English meetings: what the evidence shows.

Short answer. Yes, and English has more evidence behind it than any other language in this series, but Indian accents cost accuracy. English is 65% of the training audio in Whisper’s paper, and large-v2 scored 4.2% word error rate on read English (FLEURS). The Svarah paper tested 9.6 hours of English from 117 speakers across India: Whisper large scored 7.2%, against 1.6% and 2.7% on two native-speaker sets, and 11.2% on everyday-use prompts full of brand names, bank names and document IDs. Results depend on the speaker’s mother tongue, from 4.5% for Maithili speakers to 11.6% for Bodo speakers. The 2011 Census counts only 2,59,678 people with English as a mother tongue, because almost all English use in India is as a second or later language. ANVE’s meeting-minutes maker handles English in the development build. Coverage and figures are as of October 2026.

What open models cover for Indian English

Task	Models	What the source says
Speech to text: Whisper overall	Whisper large-v2, large-v3, turbo	English is 438,218 of the 680,000 hours in Whisper’s paper (65%). Large-v2 scored 4.2% word error rate on FLEURS English. The large-v3 card claims 10% to 20% fewer errors than large-v2 across languages, with no English-specific figure.
Speech to text: Indian accents	Whisper base, medium, large	Svarah paper (9.6 hours, 117 speakers, 65 locations in 19 states): base scored 13.6%, medium 8.3% and large 7.2%, against 2.9%, 1.7% and 1.6% for the same models on native-speaker read speech (SAA L1). Large scored 2.7% on LibriSpeech clean. The commercial Google and Azure systems scored 20.7% to 30.0% on Svarah in 2023.
Where the errors fall	Whisper large	By speaker’s mother tongue, from 4.5% (Maithili) and 5.3% (Hindi, Tamil) to 9.8% (Nepali), 10.1% (Assamese) and 11.6% (Bodo). By kind of speech: read 6.2%, extempore 7.4%, and 11.2% for everyday-use prompts with brand names, bank names, food items and document IDs.
Chat, summaries, drafting	Llama 3.3 and 4, Qwen3, Gemma, gpt-oss	Llama 3.3, Llama 4 and Qwen3 all list English. None of the cards read for this page reports a test on Indian-English text, so any claim about house style, spelling or Indian business phrasing is for you to check.

A second language, the Census and the Constitution

The 2011 Census counted 2,59,678 people with English as their mother tongue, 0.02% of India, because it asks for the language the person’s mother spoke to them in childhood. The Svarah authors estimate that about 10%, or 130 million, of Indians speak English, which is why accent is the real question. English is not in the Constitution’s Eighth Schedule; the Home Ministry lists it among 38 languages for which inclusion has been demanded. Article 343 of the Constitution kept English in official use for 15 years from the Constitution’s start and lets Parliament provide by law for its use after that. This is general information, not legal advice.

The Svarah paper asked speakers to say things like “Pay 1000 rupees from my SBI bank number … to UPI ID …” and to add named grocery items to a shopping cart. Those everyday-use prompts were the hardest: 11.2% for Whisper large, against 6.2% for read speech. The authors put it down to entities such as brand names, bank names, food items and document IDs. Office speech is full of them.

Whisper’s README lists about 1 GB of video memory for the base model, 5 GB for medium and 10 GB for large, and the paper’s base model nearly doubled the error rate on Indian accents. Model size is a real trade-off.

English with Hindi or regional words inside it is a separate test; see the Hinglish page. No source reports a result for it.

What to test before you rely on it

Record the speakers you actually have. The spread from 4.5% to 11.6% in the paper is by mother tongue, so a mixed office will land somewhere in between and individuals will vary.
Say the kinds of thing the paper found hardest, using your own examples: bank and brand names, document IDs such as PO and invoice numbers, GSTIN, PAN and IFSC codes, and UPI IDs.
Say rupee amounts the Indian way, “two lakh fifty thousand”, and check they appear in the format your documents use, such as ₹2,50,000.
Test the model size you can run, not the largest one in a paper. The paper’s base model scored 13.6% against 7.2% for large.
Check British or American spelling in the output against your house style, and whether Indian terms such as “prepone” are kept or “corrected”.

Where ANVE stands

ANVE’s meeting-minutes maker handles English transcripts today, in the development build. Every action item must quote the transcript line it came from, and a person must confirm each owner before the item is approved; each confirmation is logged. ANVE has published no accuracy or speed figures for Indian-accented English or for its hardware, so the numbers above describe the open models, not ANVE.

Speech models improve, and an ANVE box gets them as updates: ANVEAI installs updates on site today, signed packages applied offline by USB are not on ANVE yet, and ANVE never updates itself over the internet. The security page has the details. ANVE is pre-release: no sealed unit exists and there are no customers yet.

Questions people ask

How accurate is offline speech recognition on Indian-accented English?: On the Svarah benchmark Whisper large scored 7.2% word error rate against 1.6% on a native-speaker read set, and 11.2% on everyday prompts full of brand names and IDs. The paper dates from 2023 and tested the original Whisper base, medium and large models.
Does the speaker’s mother tongue change accuracy?: Yes. In the Svarah paper Whisper large scored 4.5% on Maithili speakers’ English and 5.3% on Hindi speakers’, but 9.8% on Nepali, 10.1% on Assamese and 11.6% on Bodo speakers’. Test your own speakers.
Should I use a bigger Whisper model for Indian accents?: Bigger did better in the paper: base 13.6%, medium 8.3%, large 7.2%. Whisper’s README lists about 1 GB, 5 GB and 10 GB of video memory for those sizes, so match the model to the hardware you have.
Is English counted in the Indian Census?: Yes, as a mother tongue: 2,59,678 people in 2011, or 0.02%. Most English use in India is as a second or later language, which the mother-tongue table does not capture.

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 Indian English?Accents, names, rupee amounts and Hindi words in English meetings: what the evidence shows.