Why Hinglish Breaks Every Speech-to-Text App (I Measured It)

I build voice products for a living, and this week one of them mangled its own name three different ways in a single evening. The same night, a Hinglish sentence I spoke came back written in Devanagari — with the English words transliterated INTO Devanagari. "Super cool" became "वपर पूल".

So I stopped shipping and started measuring. This post is what a night of controlled tests against Whisper (Groq's whisper-large-v3-turbo) actually showed about Hinglish — and why most transcription apps get it structurally wrong, not accidentally wrong.

The language setting is a translation switch in disguise

Every Whisper-family API has a language parameter, and every integration guide tells you to set it. Here is what it actually did in my tests:

Read that first row again. Setting language: "en" doesn't force English *transcription* — it forces English *output*. If the speaker used Hindi, the model quietly translates. For a journal, a meeting note, or a legal draft, that means the words on the page are not the words you said. No error, no warning.

For a code-switching speaker, both fixed settings are wrong by definition. "Meeting ko reschedule kar do to Thursday afternoon" is one sentence, spoken in one breath. It is not a Hindi sentence and it is not an English sentence, and an engine forced to pick one will damage it either way.

Auto-detect alone doesn't save you

With language unset, the transcription stopped translating — but the decoder still had no idea what my product names were, and it invented homophones every single time. That's not a bug in Whisper; the default decoder has simply never heard of your product, your colleagues, or your city's street names. It reaches for the nearest word it knows.

The documented fix is the prompt parameter. What the docs don't tell you is how it actually behaves.

The prompt is a decoder prior, not an instruction

Whisper doesn't *read* your prompt like an assistant. It conditions on it as if it were the transcript that came immediately before your audio. That one fact explains all four behaviors I measured:

• Names must appear inside sentences, not in a comma-list. A bare list of product names still produced mangled homophones. The same names used inside a natural example sentence came back spelled correctly every time.
• The script of the prompt controls the script of the output. A romanised prompt held my Hinglish in Latin script. Adding one Devanagari sentence to the prompt preserved pure-Hindi passages in Devanagari. You are not telling the model what to do — you are showing it what "the conversation so far" looks like.
• Recency wins. Whatever sits at the END of the prompt pulls hardest on the output. Put your most common speech pattern last.
• A generic example can collide with real speech. One of my exemplar sentences was too close to something I actually said — and it dragged my real Hindi sentence into Latin script. Your examples need to establish a pattern without impersonating the user.

I also tried plain instructions — "transcribe in the script spoken", that kind of thing. They did almost nothing. Examples did everything. If you take one line away from this post: Whisper listens to what your prompt does, not what it says.

Why this is the whole product

None of the above is exotic knowledge. It's a night of tests anyone could run. But general-purpose transcription apps can't act on it, because their defaults are built for the average English speaker — one language per utterance, one script, dictionary words.

Hinglish is not an edge case in India. It is the default register of how a few hundred million people actually speak — Hindi and English woven mid-sentence, sometimes mid-phrase. An engine tuned for single-language speech treats half of every sentence as noise to be corrected.

That's the gap IndianWhisper is built around: code-switching as the primary case, not the exception. Indian English accents, Hinglish switching, and the script coming out the way you'd write it yourself — with the accuracy work concentrated exactly where general tools stumble.

If you dictate the way you actually talk — "kal ka standup move kar do, and send the summary to Rahul" — download the Mac app and try one real sentence. That first minute is the entire pitch.

— Dhruv