On this page
A common request is an agent that follows the caller from Hindi to English and back. zlicx doesn't do that today, so here's what it does do and how to work with it.
How language works now
Each agent has one configured language. The console currently offers Hindi, English and English (India). The speech model, voice and prompt are all set up for that language, and the agent stays in it for the call.
We've chosen not to imply more than this. Automatic switching needs detection that is fast and reliable from a few words on a phone line, plus a voice that sounds right in both languages. A wrong switch in the middle of a call is worse than none. More Indian languages are planned, and we'll name them when they're available.
Serving Hindi and English callers
For now, make separate agents and choose between them before the call starts:
- Inbound, by number. Put Hindi and English agents on different numbers, for example by region or campaign.
- Outbound, by what you know. If your CRM has a language preference for a customer, use it to pick the agent when you add them to a batch.
- In your existing IVR. If you have a menu already, let "press 1 for Hindi" route to the right number.
Keep the two prompts in step. When a policy changes, update both, and keep a list of what each agent is allowed to do.
Mixed speech is still normal
Few Indian callers speak a "pure" language. A Hindi speaker will say "order cancel kar do" or read out an address in English. So test this even in a single-language agent.
A speech model built for Indian languages, such as Sarvam Saaras, is one option to try on our engine. Run the same recorded phrases through each candidate and compare the results on:
- Numbers, amounts and dates
- Names of people and places
- English product words inside Hindi sentences
- Short answers such as "haan", "nahi" and "theek hai"
Effective language support depends on the providers you choose, so judge it on your own callers' speech.
Write for the ear
Whatever the language, a prompt written for reading will sound wrong when spoken:
- Tell the model to write amounts the way people say them. Text-to-speech may read "₹1,299" differently from what you expect.
- Have it read phone numbers and order IDs in small groups.
- Prefer everyday spoken Hindi over formal written Hindi. Most callers don't talk like a government notice.
- Keep sentences short, which also helps the agent respond sooner.
Test with people, not only with text
Listening reveals what reading doesn't. Before going live, have colleagues from your target region phone the agent, ideally on mobile networks and from noisy places. The call test kit includes language cases.
Different speech models and voices have different costs, and pricing explains how these are billed. The AI receptionist page shows how this applies to inbound lines. If your callers need a language that isn't listed, let us know, because it helps us decide what comes next.
Set up your inbound line by language.
Route callers to the right agent before the call starts.