How text-to-speech works on a call
A modern TTS model reads a sentence, works out how it should sound (which words to stress, where to pause, how to say each name) and produces audio a piece at a time. On a call, that audio streams to the caller as it's made, so the agent starts speaking before the whole reply is finished.
Indian calls test a voice in ways a demo clip doesn't:
- Amounts. ₹1,25,000 should come out as "ek lakh pachchees hazaar" or "one lakh twenty-five thousand", never as a string of digits.
- Dates and times. "5 tareekh" and "10:30 am" need to sound the way people say them.
- Mixed sentences. In "Aapka order dispatch ho gaya hai", the English words should sit naturally inside the Hindi.
- Names. Customer names, brands and localities often need a pronunciation note.
Why it matters
The voice is the first thing a caller judges. A flat or stumbling voice gives people a reason to hang up before the agent has said why it called. A clear, warm voice that gets names and numbers right gives the conversation a fair start.
How Kallix uses it
In the agent studio you pick from a library of named Indian and international voices, each with samples, and set a primary and a secondary language. The greeting is ready before the caller picks up. Kallix voices are human-like, and every agent says it's an AI from the first line.