Customers hang up on voices that mispronounce their city, mash English digits into a Tamil sentence, or take a full second to notice they were interrupted. Those are product bugs, not “model limitations” you apologize for on a pricing page.

Kupe ships languages people actually speak, in the script they actually read, with interruptions and numerals treated as first-class. Cloning is available when you have rights to the donor audio. A cloned brand voice that cannot be interrupted is still a bad agent. It is just a bad agent that sounds like you.

The same voice should work on the phone and in your in-app experience. Split stacks are how you get two personalities and a support queue full of “the bot sounds different on WhatsApp.” One library. One agent. Every surface.

Listen to your own calls. If you would not leave your mother on that line, do not put customers on it. This is not a slogan. It is a release checklist. Founders skip it because the waveform looks expensive. Waveforms are not QA.

India is not a locale pack. Endpointing is different. People mix English and vernacular in one breath. Dates, rupees, and account numbers have a spoken form that is not the written form. If your vendor treats this as “we’ll fine-tune later,” later never comes.

A good voice is brief. It leaves space. It does not perform. Contact-center speech is not a podcast. The customer wants the slot, the balance, the ticket number. Give them that, in a voice that belongs on their street, then stop talking.

Cloning without rights is not a feature. It is a lawsuit. We will not help you steal a voice. We will help you clone one you own, and then we will make sure it still barges-in cleanly, because a pretty voice that talks over people is still an IVR.

After-hours clinics, collections floors, insurance renewals — they all need a voice that does not panic when the customer is noisy, in a vehicle, or switching languages. Lab audio is quiet. India is not. Test in the noise you will actually get.

Pick one voice per campaign. Do not rotate for fun. People build trust with a person, even a synthetic one. If Priya called yesterday, Priya should call today, unless Priya transferred to a human.

We keep a library so teams can start this afternoon. Swap later if you clone. The agent does not care which voice you pick, as long as it can still think out loud, look things up, and hand off. That is the stack. The voice is how it arrives.

If this page is the first time you have thought about numerals and barge-in, listen to ten of your current IVR calls before you buy anything. You already know what “almost human” sounds like. It sounds like hangups.