Voice to text for doctors India: what does it really hear? Locale models, Hinglish, brand names, and what a vendor's 'Hindi support' claim usually means.

Voice to text for doctors in India means a speech recognition model running in one locale, usually English (India) or Hindi (India), on a short audio clip. It understands the language it was set to, copes partly with words borrowed from another, and misses or mangles anything it has never seen written: brand names, salts, dose numbers run together with units. "Hindi support" on a vendor's page can mean three different things, and only one of them is Hindi speech recognition.

How speech recognition picks its words

A speech-to-text service turns audio into text with a model trained on one language in one region. Each pairing has a code called a locale: Microsoft's Azure AI Speech lists `en-IN` as English (India) and `hi-IN` as Hindi (India), each a separate row in its language table with its own real-time, fast and batch transcription support. The locale is sent with the audio. If the software sends `en-IN`, the service treats every sound as Indian English and chooses the English words that best fit.

This is why Hinglish is only half handled. A sentence such as "Dolo 650 de do, teen din tak" sent to an English (India) model gets "Dolo 650" and the numbers right, then produces whatever English words sound closest to the Hindi part. Microsoft's documentation for its fast transcription API says the language identification feature "is designed to identify one main language locale per audio file", and points to a separate multilingual mode for mixed content. A clinic product usually picks English (India), because medicine names and doses are spoken in English even inside a Hindi sentence.

What trips it up in a clinic

Audio from a consulting room has problems a meeting recording does not.

The brand name problem matters for a second reason. Regulation 1.5 of the Indian Medical Council (Professional Conduct, Etiquette and Ethics) Regulations, 2002, as substituted in 2016, says every physician "should prescribe drugs with generic names legibly and preferably in capital letters". A voice feature that only hears the brand is one step short of the regulation; the generic name still has to come from somewhere, and someone has to check it. Our post on generic name prescription rules covers the regulation in full.

Why the transcript must be editable before anything happens

Every error in the table is silent. The model does not say "I am not sure"; it returns its best guess as clean text with punctuation and capitals. Microsoft describes the output as "the best inference or prediction in text format of what was spoken". If the software passes that guess straight into a medicine list or a prescription, the guess becomes a record. The safe design shows the text to the person who spoke it, lets them correct it, and does nothing until they confirm. A feature that goes from audio to saved record with no editable step is not ready for a clinic, whatever its accuracy figure.

Transcription, extraction and generation are three different features

Vendors use "AI voice" for three things with very different risk. Ask which one is on offer.

The first two change how fast a doctor types. The third changes what the record says, which is a different risk class with its own questions: who checked the draft, what the model was told about the patient, and where that data went. Our post on what a doctor checks in AI prescription software goes through those checks; the glossary entry on AI medical scribes describes the ambient note-taking products that sit in the third row.

What "Hindi support" usually means

When a product page says it supports Hindi, ask which of these it means.

• The interface is in Hindi. Menus, buttons and patient notices are translated; nothing is said about speech.
• Output in Devanagari. A prescription or patient message can be printed in Hindi script, often by translating English text. Still nothing about speech.
• Hindi speech recognition. The software sends audio to a Hindi (India) model, `hi-IN` or its equivalent, and returns Hindi text. Only this one lets a doctor dictate in Hindi, and it meets every problem in the table above, plus English medicine names inside a Hindi sentence.

A product can honestly offer the first two and none of the third. The problem is a page that lets you assume it.

Questions to ask the vendor

1. Which locale does the voice feature send: English (India), Hindi (India), or something else? Ask for the code.
2. Which languages does it recognise, as opposed to display? Say one Hinglish sentence in the demo and read the text back.
3. Is the audio stored? By whom, for how long, and can the clinic have it deleted? Microsoft's data page for its speech service states that for real-time speech to text and fast transcription it "does not retain or store the data provided by customers"; the vendor's own storage is a separate question.
4. Who processes the audio, and in which region? Under the DPDP Act the clinic is the Data Fiduciary and the vendor and its cloud provider are processors; our DPDP guide for clinics explains what that puts on the clinic.
5. What else is sent with the audio: the patient's name, the clinic name, earlier notes? The answer should be "only the clip".
6. Can the feature be switched off for the clinic, and does the rest of the product work without it?
7. Is the text editable before anything is saved, and is anything saved automatically if the person closes the screen?

How Vitels does it, as a worked example

Vitels uses voice for one narrow job: adding medicines to the clinic's list. A doctor or clinic admin records a voice note of up to 60 seconds, or types the names. Azure AI Speech transcribes the clip in the English (India) locale, `en-IN`; it copes with Hinglish phrasing but it is not Hindi speech recognition. Claude, Anthropic's model, keeps only the medicine names from the transcript, and that text is shown for editing. "Generate" then returns one card per name with a draft composition, type and company to accept, edit, reject or skip. Nothing is saved until someone accepts a card, the Rx or OTC status is always chosen by the person from the pack, and Vitels does not store the audio. Only the clip goes to Azure and only the medicine names go to Anthropic; no patient data, clinic name or staff details are sent.

Vitels does not dictate prescriptions or consultation notes, does not generate clinical content, and has no AI scribe. The feature description is on Add medicines with AI, the data flow on AI data handling, and the term itself in the glossary under voice to text for doctors.