Mixed-language callers do not wait for your architecture diagram. They switch languages because that is how they speak. The agent should keep up.
The Hindi-English Stack
| Layer | Recommended starting point | Why |
|---|---|---|
| Language | Hindi-English where available. | This exposes the right language-specific controls, including Hinglish Map. |
| Transcriber | Soniox STT RT v4. | DialNexa sends English and Hindi hints so the transcriber expects code switching. |
| Voice | Cartesia or an ElevenLabs voice tested on your phrases. | The right text can still fail if the voice sounds formal or mispronounces local terms. |
| LLM model | A model that follows language instructions and functions reliably. | The model must know when to answer in English, Hindi, or Hinglish. |
| Prompt | Include real mixed-language examples. | The model learns the expected style from the examples you provide. |
Language Selection Details
The dashboard sorts English first, Hindi second, Indian languages next, and then other languages alphabetically. That ordering keeps common India-focused Voice AI setups close to the top without hiding other supported languages.| Detail | Why it matters |
|---|---|
| Voice-specific languages are preferred where available. | The language list reflects what the selected voice can actually speak. |
| Voice model language lookup is used as a fallback. | Older voice metadata can still produce a useful language list. |
| Some unavailable languages are hidden. | Users should not be offered languages that are not ready for selection. |
| Flux is English only. | Switching to Flux can force the language back to English where available. |
Soniox Versus Deepgram For Multilingual Calls
| Need | Start with |
|---|---|
| Hinglish or Hindi-English code switching | Soniox STT RT v4. |
| Mostly English with occasional Indian names or places | Deepgram Flux (English only) or Soniox. Compare on real recordings. |
| English-only fast turn-taking | Deepgram Flux (English only). |
| Non-English calls with a stable single language | Soniox. Test with real caller audio before scaling. |
Hinglish Map
Hinglish Map is for phrasing, not translation of the whole agent. Use it to replace formal Hindi terms with natural mixed-language words that your callers actually use. When Hinglish Map entries are configured, the runtime tracks the caller’s current language style on each turn. It can identify English, Devanagari Hindi, Romanized Hindi, and mixed Hinglish, then adds a language instruction before the LLM responds. After the LLM responds, DialNexa can soften formal Hindi terms with the built-in casual Hindi map plus your workspace entries before audio is synthesized.| Formal phrase problem | Better approach |
|---|---|
| Agent sounds too official for a casual reminder call. | Map repeated formal terms to conversational Hinglish. |
| Caller uses English nouns inside Hindi sentences. | Add prompt examples that show the same mixed style. |
| Voice over-pronounces English product names inside Hindi sentences. | Test another voice or voice model before adding confusing prompt hacks. |
Language tracking helps the agent choose the response style. It does not change the selected TTS voice. Use a voice that can speak the language and script you expect callers to use.
Write Prompts For Mixed Language
Name the expected language behavior
Tell the agent whether it should start in Hinglish, mirror the caller, or stay mostly English with Hindi support.
Include sample caller phrases
Add phrases that callers actually say, including short replies, local terms, city names, and English product words.
Include sample agent replies
Show the tone you want. One good example can prevent twenty awkward formal sentences.
Review Checklist
| Check | What to inspect |
|---|---|
| Greeting | Does the first sentence sound natural in the expected Hindi-English mix? |
| Names and places | Do Indian names, cities, and localities transcribe correctly? |
| Numbers and amounts | Can the voice read rupees, dates, times, and phone numbers clearly? |
| Turn-taking | Does the agent wait through mixed-language pauses? |
| Transcript | Does post-call analysis receive the same facts the caller said? |
| Voice tone | Does the selected voice sound conversational rather than translated? |
Common Multilingual Mistakes
Prompt is English only
Prompt is English only
The model may answer in English even when the caller expects Hinglish. Add explicit language instructions and examples.
Voice sounds formal
Voice sounds formal
Use Hinglish Map and test another voice or model. The selected speaker matters as much as the text.
Variables are not language-aware
Variables are not language-aware
A city, plan, date, or payment amount injected through variables may need wording around it so it sounds natural.
Post-call fields depend on translated wording
Post-call fields depend on translated wording
Field descriptions should describe the business meaning, not the exact language used during the call.
Related Reading
Supported Transcribers
Compare Soniox and Deepgram options.
Speech Settings
Use Hinglish Map and Response Eagerness.
Text To Speech
Choose a voice that fits mixed-language calls.
Transcripts And Recordings
Audit mixed-language evidence.