Can AI transcribe Hinglish correctly?
Mostly right on the words, not yet on the spelling.

Whisper, the open speech model that many Hinglish fine-tunes start from, was trained to predict one language label at the start of each 30-second window. Its authors also left out of the transcription training data any pair in which the spoken language did not match the language of the transcript; English transcripts were kept only as translation examples. That recipe was not designed around mixed-language speech.

Then there is spelling. Hinglish has no standard written form: the same sentence can be typed in Roman letters, in Devanagari, or in a mix. The team behind the Vaani Benchmark notes that current test sets often lack enough code-switched speech or use inconsistent script conventions, so they transcribe code-switched words in both scripts. A model that writes the word for a meeting as मीटिंग may be exactly right and still be marked wrong against a Roman-script reference.

Hinglish figures come from different test sets with different spelling rules, so read the table as a set of clues, not a league table.

• The spelling of English terms and brand names. The same word may appear in two scripts in one transcript.
• Names of people, places and products, which any speech model tends to guess by sound.
• Numbers said in a mix of Hindi and English, such as pachas hazaar, fifty thousand or 50K. Agree how amounts are written, and check every figure against the audio.
• Sentences where the language switches mid-clause, which is where the language-label problem shows up.
• Lines that were never said or are repeated. The Whisper model card warns about both, especially around silence.
• Any search you intend to run later. Roman-script and Devanagari versions of the same word will not match unless your search treats them as one.

• Decide the house style before you test: Hindi in Devanagari and English in Roman letters, everything in Roman letters, or whatever your readers prefer. Put it on one page with six example sentences.
• Pick five real recordings, correct a transcript of each by hand in that style, and keep them as your test set.
• Try a general model such as Whisper large-v3 and one or two Hinglish-aware models, and compare errors by type: names, numbers, English words and Hindi words.
• Use the prompt where the tool allows it. OpenAI’s speech guide says Whisper takes a prompt of up to 224 tokens, which helps with names, acronyms and punctuation, though Whisper does not follow instructions the way a chat model does.
• Keep clips short where you can. Tara’s card says it was evaluated on clips of up to 30 seconds and that longer audio should be split.
• Keep audio and transcripts on the office server, and repeat the test whenever you change the model, the microphone or the room.

ANVE’s meeting minutes maker works with transcripts in Hindi, English and Hinglish. Every action item must quote the transcript line it came from, so a reader can check a mixed-language sentence against what was said, and a person confirms the owner before the item is approved. Each confirmation is logged.

This runs today in ANVE’s development build. No sealed unit exists yet, and ANVE publishes no accuracy figures for Hinglish of its own.

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.