When voices can be cloned, even real people face doubt. Discover OriginStory’s mission to make trust in voice verifiable, giving people evidence that they actually spoke.
For most of the history of phone calls and recordings, hearing a familiar voice was good reason to believe that person had spoken. A lot was built on that assumption: wire authorizations given over the phone, hires made virtually, records bought because a specific artist sang on them.
That assumption no longer holds. The quality of synthetic voice generation now casts doubt on the authenticity of remote communication. This greatly erodes the trust we have taken for granted. Our mission is to make trust in voice verifiable.
Our mission is to make trust in voice verifiable.
Doubt follows the real person
In the first season of the Shell Game podcast, Evan Ratliff sent an AI version of himself to talk to his friend Keegan. Keegan suspected he was speaking to a bot. But when the real Evan called afterward, he wasn’t entirely sure that voice was real either. “Ten percent of uncertainty, it turns out, is a lot of uncertainty,” Keegan said. The clone didn’t have to fool him completely to make him doubt the real person afterward.
What evidence could Evan offer that it was really him speaking? Another recording would face the same question. So would another call. For us, that is the harder problem. Catching an impersonation doesn’t necessarily restore confidence in the next conversation. A candidate answering an interview question or an artist releasing a vocal recording may have done exactly what they claim, yet struggle to demonstrate it.
We want people to have a better answer than “you’ll have to take my word for it.”
A microphone can record a loudspeaker
The checks available today answer parts of the question. Account verification can establish who logged in, but it does not establish who is speaking. A verified person can arrange for a synthetic voice to speak on their behalf—like Evan did.
AI detectors examine audio for signs of generation. Their results can be useful, but they remain judgments about the recording. In our own experiments with a commercial detector, we found that transformations to synthetic audio could change its classification to authentic. We’ll share the methods, results, and limitations in a future post.
Signed recordings address another part of the problem. The Coalition for Content Provenance and Authenticity (C2PA) develops standards for attaching verifiable information about digital media’s source and history. But a signature from a recording device does not, by itself, establish what made the sound. Put a loudspeaker in front of a microphone and play a synthetic voice. The microphone can record it faithfully and sign the result. The recording has valid provenance, but the speech still came from a model.
This is where our approach begins. To establish that a person produced the voice, we need to measure something beyond the sound itself.
Measuring speech at the source
Human speech involves physical activity in the body—lungs expanding and contracting, heart beating, larynx vibrating. Our implementation puts a small, low-power sensor beside the microphone. The sensor measures physical activity associated with speech and checks that it matches the captured audio.
The check runs on the device, and the result is signed as the audio is recorded. A more convincing voice deepfake doesn’t reproduce the physical activity we measure. The base claim of the core technology is narrow: a live human produced this voice. We call this Verified Human. Where identity matters, that evidence can be combined with a voice match and a registered device to support a stronger claim: this specific person is speaking. We call this Trusted Human.
We want someone seeing a verification result to understand exactly what they can —and cannot—conclude from it.
Building trust into the devices we use
We shouldn’t have to enter every conversation wondering whether there’s a person on the other end. Or become experts at spotting the pauses and intonations of a synthetic voice. As those tells disappear, asking people to listen more carefully is asking them to do an increasingly impossible job.
Our bet is that verification belongs in the devices and services we already use to communicate.
We’re working toward a world where a headset, phone, or other recording device can capture speech along with evidence of its human origin. That evidence can travel with the audio in existing formats such as C2PA, so the services receiving it have something to check beyond how convincing it sounds.
In a live call, that means verification alongside the speaker’s name. In a recording, it means evidence of a human performance or an official statement that remains available after publication. For a system acting on spoken instructions, it means another basis for deciding whether to proceed or ask for confirmation.
These are different uses of the same capability: giving people a reason to trust where a voice came from. The doubt Keegan described should be something our tools help us resolve, rather than something we carry into every conversation. Device makers can provide the evidence at the source; communication platforms and other services can make it useful wherever the audio goes.
Our aim is to be the layer they build it on.
Where we are
We have a working sensor, a Zoom integration that shows a Trusted Human badge during a call with the camera on or off, and a deployment with the Arizona Secretary of State authenticating official election communications.
We’re also showcasing our technology as one of the Wow demos at Infineon’s OktoberTech Silicon Valley on October 22 at the Computer History Museum. If you’re attending, come see it in action and talk with us about bringing voice verification into the devices people already use.
There is still a lot to learn. We are running pilots across recorded media, live conversations, and spoken commands to machines. We want to understand where this changes what people are willing to do, how it performs in everyday use, and where the benefit is worth deploying hardware. Those are questions we want to answer by putting the system into use. We’ll use this blog to examine where trust in voice is breaking down, test the checks available today, and share what we learn from building and deploying our own approach. Next week, we’ll look at what changes when a voice can set a machine in motion, and what it takes to keep people in control as AI systems act in the physical world.
As voices become easier to generate, we want people to retain a simple ability: to establish that they actually spoke.
