# Hinglish STT, Acronyms and Keyword Checks

> On Hindi-English calls, speech-to-text can write English acronyms in Devanagari and break keyword-based qualification. Why, and how to design around it.

Source: https://edesy.in/blog/hinglish-stt-acronyms-keyword-qualification-2026

**If your qualification depends on hearing a specific English word, check how it is written in the
transcript.** In one deployment, a CBSE exam-prep book publisher's agent looked for designations
like PRT, TGT and PGT that genuine teachers mention unprompted. On a Hindi-English benchmark, the
speech-to-text transcribed some of those acronyms in Devanagari. A check for the English string
would have marked real teachers as unqualified. This post explains the problem and how to design
around it.

> **Build a Hinglish agent that qualifies correctly:** [Start free on Edesy](https://voice-agent.edesy.in/?utm_source=edesy.in&utm_medium=blog&utm_campaign=hinglish-stt-acronyms-keyword-qualification-2026) — Rs 50 free credit, no demo needed.

## What happens

A teacher says: "Main TGT hoon, Class 9 aur 10 ko science padhata hoon."

Depending on the language setting, the transcript may read:

- "Main TGT hoon..." (Latin script), or
- "मैं टीजीटी हूं..." (Devanagari), or a mix.

Both are correct transcriptions of what was said. The problem is anything downstream that expects
one spelling.

## Where it breaks

- **Keyword rules in post-call extraction.** "Qualified if the transcript contains TGT, PGT or
  B.Ed" silently fails.
- **The agent's own logic.** An instruction like "if the caller says TGT, skip the class question"
  may not fire.
- **Reports and searches.** Searching transcripts for "CTET" misses "सीटीईटी".

The same applies to other common English terms on Indian calls: EMI, KYC, OTP, GST, COD, PIN code,
brand and product names.

## How to design around it

1. **Judge meaning, not strings.** Ask the extraction step "Did the person mention a teaching
   designation or qualification (in any script)? Which one?" instead of matching exact text. LLMs
   handle Devanagari and Latin spellings of the same acronym well when told to.
2. **List both spellings in instructions** for the few terms that matter: "TGT (टीजीटी), PGT
   (पीजीटी), B.Ed (बीएड)".
3. **Don't make one keyword decisive.** Use it as one signal alongside class, subject and how
   detailed the answer is.
4. **Keep the language setting that fits your callers.** Switching to English-only transcription to
   fix acronyms usually makes the Hindi parts worse.
5. **Test with real audio.** Record 20-30 real or realistic calls and read the raw transcripts
   before trusting any rule.

## A quick audit checklist

- [ ] List every word or acronym your qualification or routing depends on.
- [ ] Search a day of transcripts for each, in both scripts.
- [ ] Rewrite exact-match rules as meaning-based questions.
- [ ] Re-check after any change to speech or language settings.

## Related: how the agent *says* acronyms

This post is about hearing and writing acronyms. For how the agent pronounces numbers, acronyms and
class names, see [pronunciation of numbers and acronyms in India](https://edesy.in/blog/ai-voice-agent-pronunciation-numbers-acronyms-india-2026).
For choosing between Hinglish and pure Hindi, see
[Hinglish vs pure Hindi on AI calls](https://edesy.in/blog/hinglish-vs-pure-hindi-ai-calls-2026).

## Try it

Edesy supports 75+ languages, including 14 Indian languages, including Hindi-English mixed calls, with post-call
extraction you can write in plain language.
[Create a free account](https://voice-agent.edesy.in/?utm_source=edesy.in&utm_medium=blog&utm_campaign=hinglish-stt-acronyms-keyword-qualification-2026),
or see [how a publisher verifies teachers with AI calls](https://edesy.in/ai-voice-assistant/case-studies/publisher-teacher-verification-calls).

## Related reading

- [Stop free-sample misuse with AI calls](https://edesy.in/blog/stop-free-sample-abuse-ai-verification-calls-2026)
- [AI call QA: evaluate after the call ends](https://edesy.in/blog/evaluate-ai-calls-after-hang-up-qa-2026)
- [Verification questions on AI calls that verify](https://edesy.in/blog/verification-questions-ai-calls-dont-give-options-2026)

## Frequently Asked Questions

### Why does speech-to-text write English words in Hindi script?

On mixed Hindi-English audio, a speech-to-text model set up for Hindi may transcribe English words and acronyms in Devanagari, for example writing 'TGT' as 'टीजीटी'. The words are heard correctly but written in a different script.

### How does that break lead qualification?

If your agent or your post-call rules look for the English string 'TGT', 'B.Ed' or 'EMI', they miss the Devanagari version, so qualified people are marked unqualified.

### How do you fix keyword checks on Hinglish calls?

Match meaning, not strings. Ask the LLM to judge whether the person mentioned a designation or qualification in any script, list both spellings in instructions, and test with real recordings before relying on the result.

### Should I switch to English-only transcription?

Usually not. English-only transcription handles the Hindi parts of the call worse. Keep the language setting that fits your callers and make your checks script-aware.

### How can I catch this problem early?

Read the transcripts of your first 20-30 calls, not just the summaries. If key terms appear in Devanagari, review every rule that depends on them.