Can AI transcribe Hindi meetings offline?
Yes, on your own hardware, as a draft that a person still reads.

Once a model has been downloaded, none of the tools below needs an internet connection to transcribe. Audio and text stay on the machine that runs them. All are free to download, though some ask for contact details first, and the licence column (as listed in October 2026) matters before an office relies on one.

The most useful recent test is the Vaani Benchmark, published on arXiv in 2026 by researchers at the Indian Institute of Science and its ARTPARK technology park. It holds 5,050 clips, about 11 hours, from 1,062 speakers in 102 districts, all spontaneous speech recorded in ordinary conditions. Each system is scored strictly, against one human transcript at a time with the best of the three reported, and again leniently, by accepting any of three human spellings.

Whisper’s weaker showing on Hindi has a history. The original Whisper paper lists only 12 hours of Hindi transcription among its 117,000 hours of audio in other languages, and finds that the amount of training data for a language is very predictive of how well the model does on it. Large-v3 was trained on a larger and different mix, so treat the 12 hours as background, not as a measurement of v3. The same paper shows that model size matters: on its Common Voice 9 Hindi test the small model scored a 43.6 per cent word error rate and large-v2 scored 21.9 per cent.

• Names, amounts, dates and phone numbers. Check each against the audio, not against the draft.
• English words inside Hindi sentences. Models may write them in Devanagari, in Roman letters or inconsistently. Agree one house style first.
• Lines nobody said. The Whisper model card warns that predictions may include text that is not actually spoken, and that the model can repeat itself. Read the transcript especially where the room went quiet.
• Accents and specialist words. A 2024 accuracy study of Hindi speech (LAHAJA) found that performance fell for speakers from the North-East and South India, and on named entities and specialised terminology.
• Who spoke. Transcription gives words, not names. Speaker labels are a separate step with their own error rate.
• A transcript that looks different from a reference. In the Vaani benchmark three independent human transcribers differed from one another on 10.4 to 13.3 per cent of words, partly because spelling and word choice have several acceptable forms. A transcript can be right and still not match.
• Anything that will be quoted, signed or acted on. A person reads the whole transcript first.

• Collect five real recordings of about ten minutes each: a quiet one, a noisy one, a phone call, a meeting where Hindi and English mix, and one full of names and figures.
• Have a colleague correct a transcript of each. That corrected text is your reference.
• Run two or three candidate models, for example Whisper large-v3, IndicConformer and Qwen3-ASR, on all five and compare the word error rate and the kind of mistakes.
• Choose the lightest model that clears your bar. In its README, faster-whisper timed its small model at 1 minute 42 seconds for 13 minutes of audio on a desktop CPU, and the large-v2 model at 59 seconds on an 8 GB gaming GPU. Small models cost accuracy, as the Hindi figures above show.
• Record with one decent microphone in the middle of the table, not a phone in a pocket.
• Keep audio and transcripts on the office server, set a retention period, and delete audio once the minutes are approved.
• Repeat the five-recording test whenever you change the model, the microphone or the room.

ANVE’s meeting minutes maker works with transcripts in Hindi, English and Hinglish, on a device in the office. Every action item in the minutes must quote the transcript line it came from, and a person has to confirm its owner before the item is approved. Each confirmation is logged.

This runs today in ANVE’s development build, on a development machine. No sealed unit exists yet, and the final price is not announced. ANVE publishes no speed figures for its hardware, so use the timings above only as a guide to the open tools.

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.