When Sanas launched the first version of accent translation, Shawn Zhang did not watch the metrics from a dashboard in San Francisco. He flew to the Philippines and lived on a call center floor for a month, listening to how agents actually work, where the audio actually breaks, and what a speech product actually has to survive.
That instinct, to go live inside the problem before building anything, runs through this episode of Skywatch, Bluejay’s car podcast. Rohan sat down with Shawn, co-founder and CTO of Sanas, to talk about what it takes to build speech AI that works in the real world, and why the problem Sanas is solving is bigger than most people realize.
From Stanford to a call center floor
Shawn came out of Stanford’s ML group, where he worked under Andrew Ng on computer vision and NLP. Coming from that world, the natural assumption would be that the hardest part of building Sanas was the modeling. Shawn knew from the start it would not be.
The hardest problem was the data. There is no public corpus of real call center speech: the accents, headsets, background noise, and pacing of actual customer conversations. So Sanas’s first commercial deals were not structured around revenue at all. The company traded its product for access to real call center audio, betting that the data would compound into something no competitor could shortcut.
“We are a real-time speech AI company. We specialize in building speech AI products that transform speech to sound more understandable.”
Three products, one idea
Accent translation is what Sanas is best known for: changing a speaker’s accent in real time so they can be understood anywhere in the world, without changing their voice or what they are saying. But it is one of three products that all serve the same idea.
Speech enhancement goes beyond removing background noise; it enhances articulation and fidelity, so the speech that comes through is not just cleaner but genuinely easier to understand. And real-time language translation lets people speak across languages with as little latency as possible.
Together, the suite has over 1.2 million people using it live, most visibly in customer service operations across India, the Philippines, Latin America, Africa, North America, and Europe.
Eighteen million hours: the data moat
The number that anchors this conversation: Sanas has collected more than eighteen million hours of speech data across six hundred accents and dialects.
That scale is not a vanity metric. Real conversations are messy in ways demo environments never are, and a model that has only seen clean speech falls apart the moment it meets a crowded floor, a cheap headset, or an accent it was never trained on. In Shawn’s telling, that corpus is precisely what separates a good demo from something that actually works in production, and it is the advantage that compounds with every deployment.
The future of customer service is hybrid
A recurring theme in the episode: the future of customer service is not fully automated. AI agents make sense for handling scale and volume, but human agents stay in the loop for the escalated, nuanced cases where empathy and judgment matter most.
That is what makes Sanas’s position interesting. Its products sit underneath both sides of that hybrid world: making human agents easier to understand on every call, and making the whole conduit of speech clearer for whatever is on the other end of it.
