The voice on the other end of the line is still human.

A call center in Manila. A buyer in Ohio. Nobody asks for a supervisor.

It is 2 a.m. somewhere. An agent in Manila or Bogota or Nairobi puts on a headset, taps a small toggle on her screen and answers a call from a man in Ohio who has had a bad day with his router. He hears her warmth. He hears her words. He does not hear the thing he has been trained, by twenty years of customer service hell, to listen for first - the gap between his accent and hers. He just hears help. The call ends in seven minutes instead of seventeen. Nobody asks for a supervisor.

That edit - the part where the friction between two human accents stops mattering - is what Sanas sells. The company is six years old, headquartered above a coffee shop on Hamilton Avenue in Palo Alto, and quietly runs inside twelve of the twenty largest business process outsourcing companies on earth. If you have called anyone's customer service line this year, the odds are not bad that you have already met it without knowing.

The problem they actually saw

The standard story about call centers is a story about cost. The real story is a story about being misheard. The global contact-center industry employs roughly seventeen million people, the majority of whom are speaking English as a second, third or fourth language to customers who are mostly impatient, occasionally cruel and almost never trained to listen across an accent. A 2020 study by ResearchGate found that callers rated agents with what they called "foreign" accents lower on competence - even when the agents resolved the problem faster.

You can fix this two ways. You can demand the agents change how they speak, which is what the industry tried for forty years through accent-neutralization classes and largely failed at. Or you can fix it on the wire, somewhere between the agent's microphone and the customer's earbud, in the seventy or eighty milliseconds when nobody is paying attention.

The second option used to be science fiction. Then a Stanford undergraduate named Andres Soderi started working part-time at a call center to pay rent, got mocked for his accent one too many times, and went back to the dorm and complained.

Three immigrants and a hunch

Maxim Serebryakov, Shawn Zhang and Sharath Keshava Narayana were all first-generation immigrants when they founded Sanas in 2020. They had each, in their own way, lost arguments because of how their consonants landed. They believed - and this is the bet that took them six years to prove - that you could use neural networks to change an accent on a live call without losing the speaker's identity. Not the pitch of the voice. Not the emotion. Not the timing of a laugh. Just the accent.

It is the kind of bet that sounds either trivial or impossible, depending on whether you have ever tried to do it. The audio research community, for years, treated real-time voice conversion as a parlor trick - good for novelty TikToks, useless for anything where milliseconds mattered. Sanas's founders disagreed, politely, and then went looking for the money to be right.

They got $5.5 million in seed in 2021, $32 million in Series A from Insight Partners in 2022, and $65 million in Series B in February 2025 led by Quadrille Capital. Notably, two of the largest contact-center operators in the world - Teleperformance and Alorica - participated in that last round. They were not just investing. They were already customers.

What's actually running on the call

Sanas's platform is, technically, four things in a trench coat. There is the accent translation engine, which takes a speaker's voice and shifts its phonetic profile toward a target accent in roughly 200 milliseconds - fast enough that humans do not detect the lag, and slow enough that the neural network actually has time to think. There is the language translator, which renders one language as another in the original speaker's voice, a feature the company launched in 2024 and that is, depending on whom you ask, either the obvious next step or a genuinely uncanny leap. There is speech enhancement, which scrubs reverb, echo and the acoustic weirdness of a cheap headset in a hot room. And there is noise cancellation, which Sanas gives away free, the way Costco gives away rotisserie chicken.

The strategic logic is the same as Costco's. Once the noise canceller is in the agent's ear, the upsell is short.

What is unusual, technically, is that the system runs at the endpoint. It does not phone home to a GPU farm in Virginia. It does not need a perfect internet connection. It works in a building in Cape Town with the air conditioning broken and the fluorescents buzzing - which is most buildings, most of the time, in the industry Sanas sells into.

Who is actually paying for this

UPS uses it. DoorDash uses it. Teleperformance, which is the largest contact-center operator on the planet and now also an investor, has deployed it across multiple geographies. Alorica, the second-largest, did the same. The trade-press scoreboard now reads: twelve of the top twenty BPOs in the world, hundreds of thousands of agents, thirty-nine countries.

The customer story is mostly about money - call resolution gets faster, customer satisfaction scores climb, agent attrition drops because being mocked all day is bad for retention. But underneath the money there is something else, and it is the thing that the company's marketing keeps circling without quite landing on. Sanas's customers are, in effect, fixing a forty-year-old design flaw in the global call center: that being understood was supposed to be the agent's job, alone, unaided, against the odds.

What they actually think they are doing

The company's stated mission - "build a more understanding world" - is the kind of phrase that, in 2026, ought to make any sensible reader reach for their wallet to make sure it is still there. Mission statements are mostly varnish. But Sanas's reads slightly differently when you know that all three of its founders grew up speaking accented English in rooms where it cost them. The product is, in part, an autobiography.

Critics have not been silent. A widely-circulated 2022 piece argued that the product, taken to its logical conclusion, makes the world sound whiter - that it papers over discrimination instead of demanding the listener get over it. Sanas's answer, when pressed on this, has been twofold. First, that the agents themselves overwhelmingly opt in, because the alternative is being yelled at. Second, that the product preserves the speaker's voice and emotion - it is not erasing identity, it is unblocking comprehension. You can decide which side that argument lands on. The agents, so far, have voted with their headsets.

The bigger thing under all of this

Sanas is a contact-center company today because that is where the pain is most acute and the buyer is most obvious. It is, almost certainly, not a contact-center company forever. Real-time, voice-preserving speech-to-speech translation is the kind of capability that, once it works in a call center, tends to leak everywhere else - into telemedicine, into classrooms, into immigration interviews, into the kind of dinner with your partner's parents where everyone is pretending to follow the joke. The company has been careful, so far, not to overpromise. But the roadmap is what the roadmap is. The wire is everywhere.

What is interesting, and what makes the next five years worth watching, is that Sanas is one of the few generative-AI companies whose product gets better the less you notice it. There is no chatty interface. There is no avatar. There is just a toggle, and the conversation that follows.

So return to the call. The agent in Manila. The buyer in Ohio. Seven minutes instead of seventeen. No supervisor. Nothing remarkable, in the sense that nothing that happens on that call is the kind of thing anyone will write a headline about. Which is, of course, exactly the point. Sanas's most successful product is a call that ends with nobody noticing it was ever hard.