Our AI voice agent cold called 15,000+ businesses… here’s what we learned.
Which voices performed best in outbound calls? Turns out, it depends!
It was just a regular day in the Rime HQ, where we were debating the merits of one voice over another. We all have our favorite Rime voices, and for no other reason than “it sounds better” - which as we all know, is completely subjective. One of our founders proposed an experiment… what if we called people with various voices, a duel of sorts, to see which ones win? The voice lab was born.
Same script, same model, same day: one voice got hung up on three times as often as another.
That happened in one of our early tests in Atlanta. Our Voice AI agent called restaurants to check their closing time, and each call was randomly assigned one of five Rime voices. The only thing that changed was the voice.
The results were fascinating. With Godfrey, 15% of conversations ended in a hangup while with Ocean, it was 48%.
We build voices for a living (and a decent percentage of Rimers are linguists!), and we still couldn't predict it. Each morning before the agent started calling, our team members picked their favorites. We were running a Slack-based betting ring every day, competing for bragging rights, but it was mostly a guessing game. There are still results that we can't explain.
We published all these voices in what we’re calling Voice Duel. Each week we’re running a new experiment to see how our voices handle different scenarios and geographies.
If you're choosing a voice for a phone agent, the Voice Duel leaderboard will give you real-call evidence to help build your shortlist. So, make sure to join the community where these weekly insights will be shared first.
The question we can't answer “yet”
In our calls, many people talked to the agent the way they'd talk to any caller: they put it on hold to check an answer, they offered to book a table. In 391 of the 616 October 12 conversations that reached the agent's closing thanks, the recipient replied within two seconds. That is 63% of the people, with replies such as "you're welcome" and "bye."
Those calls didn't announce AI at the start, so we don't know how many people recognized it. The script instructed the agent to identify itself as automated when asked. One person stopped mid-call: "Are you a human or are you a robot?" The agent told them, and they still answered the rest of its questions. Another person asked "Is this AI?" during the farewell and again got the truth.
So here is the question: do people mind talking to an AI if they like how it sounds? Let's see when we will have enough evidence to answer that. We'll start by measuring which voices keep people engaged and collect useful stats with each progression.
What our early tests already showed
Pick voices, not vendors: Five voices on the same Rime model ranged from 15% to 48% hangups in a single test. A provider's name tells you little about how one of its voices will do.
Fast speech isn't the same as a fast conversation: Mist v3 produced its first audio in 52 ms, about 100 ms ahead of the others we tested. But it's worth observing how working out that the person had finished speaking and waiting for the language model each took longer than generating the speech itself.
Fun fact: One call may have had AI on both ends. The restaurant's line appears to have been answered by an AI assistant. It answered every question and offered to book a table. Two agents talking to each other... is that our future?!
I'll be in the Rime Slack Community sharing some of our findings - join me!