SI Phone Bots Now Say ‘Mm-Hmm’ and Stop When You Cut In. They Cost 3.5 Cents a Minute.

The next time a voice on the phone says “mm-hmm” while you read off your card number, you may not be talking to a person. A voice company called Fish Audio has released a model built to make SI phone agents sound like they’re actually listening, and businesses can run it for 3.5 cents a minute.

The model is called Live 3, and it went live Nov. 10. Its whole pitch is timing. Most phone bots wait for you to stop, think, then talk, and if you so much as cough mid-sentence they either barrel ahead or stop dead. Live 3 is built to listen and talk at the same time, the way two people do.

Why a pause is hard for a robot

Anyone who has yelled “REPRESENTATIVE” into a phone tree knows the problem. Fish Audio’s own launch post lays it out with a simple example. A pause after “What time do you open?” probably means you’re done talking. A pause halfway through a phone number means more is coming. A machine that treats both the same will either cut you off or sit in dead air.

Most voice agents today are built like a relay race. One piece turns your speech into text, another figures out an answer, a third turns that answer back into speech, and a separate set of rules guesses when it’s the bot’s turn. Each handoff adds a little delay and a little clumsiness.

Fish Audio says it moved that turn-taking job into the speech model itself and trained it on recordings of real human conversations, so it learned on its own when to acknowledge, wait, answer or stop.

“The fix so far has been more rules and more thresholds,” Rissa Cao, the company’s co-founder and CEO, said in the announcement. “We think a model should learn this from real conversations, the same way language models learn grammar.”

Three things it is supposed to get right

The company built its demos around ordinary moments that trip up today’s phone bots.

The first is reading out a number in chunks. Say “415… 555… 0132,” and Live 3 might answer “mm-hmm” or “okay, 415” between the pieces, then wait for the rest. That little sound tells you it heard you without grabbing the floor.

The second is telling the difference between a nod and an interruption. If the agent is listing its hours and you say “yeah” to show you’re following, it keeps going. If you say “Wait, actually, I just need Thursday,” it stops and takes the new request.

The third is letting people change their minds. The company’s example: “I’d like two pizzas… no, three. Three large. And… actually, make one of them a margherita.” Instead of jumping in after “two pizzas” and getting the order wrong, the model is supposed to sit through each correction until you’re done.

Fish Audio says the model can decide to react in as little as 186 milliseconds, about the blink of an eye, for quick sounds like “okay” and “mm-hmm,” while the full answer is still being worked out in the background.

Built for the businesses that answer your calls

This isn’t an app for your phone. It is a tool for the companies on the other end of the line: restaurants taking orders, offices booking appointments, support desks fielding complaints. Live 3 runs on Fish Audio’s agent-building platform and its API, and big companies can install it on their own servers. It speaks English only for now.

The price is what should get attention. According to the company’s product page, the platform costs 3.5 cents a minute with Live 3 included, half off the regular 7 cents. A launch review from Kie.ai says the discount runs through Dec. 31 and that phone calls add about a penny a minute, with a dedicated phone number at $1.20 a month. At those rates, an SI receptionist that never sleeps costs a few dollars a day, a fraction of an hourly wage.

Cao pushed it hard on launch day, offering to build a free custom agent for the first 1,000 people who replied with a use case. Her post passed 1.7 million views. A partner called SkillNet announced a “Voice Operator Agents” skill built on Fish Audio that it says can run a business’s phone line on its own.

Fish Audio isn’t a garage project, either. The company announced a $52 million seed round in October along with its Drama 3 voice model, and says it has more than 15 million users. Drama 3’s weights were released for anyone to download. Live 3 is closed.

The part nobody has checked yet

For all the slick demos, there is no scorecard. MarkTechPost noted that Fish Audio “publishes no benchmark scores” for Live 3, and that the 186-millisecond figure is a best case for short reactions, not how fast a full answer arrives. Outside testers have not yet put it through an independent leaderboard. Until they do, the claims are the company’s.

There is a bigger question for everybody who picks up a phone. Fish Audio’s library includes more than 2 million voices, including cloned ones, according to MarkTechPost. A bot that sounds bored, polite and human, says “mm-hmm” at the right moment and never trips over an interruption is great for ordering pizza. In the wrong hands it is also a better scam caller. The FCC already ruled in 2024 that robocalls using SI-generated voices fall under the same federal law as other robocalls, which means they need your consent first. Enforcing that against a voice you can’t tell from a person is another matter.

So here is where this leaves the rest of us. The robotic phone tree is on its way out, and the replacement is going to sound a lot like the person you thought you were talking to. If a call ever feels too smooth while someone is asking for your bank details, hang up and call back on a number you trust.