engineering

Why most voice agents still sound like a 2009 IVR

Kupe Team · June 12, 2026

Ask anyone who has run a contact center: the fastest way to lose a customer is dead air. They asked a simple question. The system is looking something up. The line goes quiet. They think the call dropped, or worse, that they reached a fake number.

Most voice AI demos hide this. The script is short. There is no CRM. Nobody interrupts. The founder is on stage with a headset and a friendly crowd. In production the opposite happens. The caller talks over the agent. The agent needs two seconds to fetch a balance. An English “um” gets dropped into a Tamil conversation. Trust evaporates in a breath.

We spent a year listening to those hangups. Not the ones where the product was wrong — the ones where the product was fine and the line still felt empty. People will wait if they hear a person thinking. They will not wait for a spinner they cannot see.

That gap — the beat between “the customer stopped talking” and “the answer is ready” — is where voice products go to die. Chatbots can show a typing indicator. Phone calls cannot. WhatsApp can show a clock. A live number cannot. If you do not fill that beat, you have built a voicemail with better copy.

ThinkSpark is how we fill it. It is a small model, not a canned soundboard. It predicts a short, language-appropriate thinking sound in a few milliseconds, on CPU, next to the live call. Hindi does not get an English filler. Tamil does not get a cartoon beep. The caller hears presence.

That sounds like a research footnote. It is actually the difference between a collections agent that recovers money and one that generates complaints. It is the difference between a clinic night line that books Thursday and one that loses the patient to a competitor who picked up.

If you are evaluating voice platforms, put a real lookup in the middle of the script and interrupt the agent. Ask for a balance. Ask for a slot. Talk over the greeting. If the line dies, you do not have a production stack. You have a playground with a phone number glued on.

We would rather ship the unglamorous millisecond than another landing-page waveform. Contact centers do not buy waveforms. They buy conversations that hold. A waveform impresses a board deck. A thinking sound keeps a borrower on the line while the ledger loads.

ThinkSpark is on by default for Kupe agents. The weights are public on Hugging Face if you want to inspect them. The hub is where you hear it on a live number, in the language your customers actually speak.

The unfashionable truth: most of voice AI is not the model that talks. It is the model that does not leave you alone in the dark while the important model thinks. We built that first, because that is what the night shift asked for.

When teams say “your agent feels human,” they rarely mean the vocabulary. They mean they were not abandoned. Presence is a product. Silence is a bug. We treat it that way.

If this sounds small, listen to twenty production calls from any vendor that skipped it. You will hear the hangups. You will hear the “hello? hello?” You will hear why we refused to ship without it.

Why most voice agents still sound like a 2009 IVR · Kupe