California-based startup Subtle Computing is tackling the issue of capturing folks’s voices in noisy environments with its personal voice isolation fashions — a know-how that might profit voice-based AI services and products.
Shopper apps utilizing voice AI are at present seeing great progress. AI Assembly notetakers like Granola, Fireflies, Fathom, and Learn AI have acquired each consumer and investor consideration. Present corporations like OpenAI, ClickUp, and Notion have built-in voice transcription options. App makers like Wispr Movement and Willow are engaged on voice dictation. Then there are {hardware} corporations like Plaud and Sandbar which might be utilizing units as a medium to transcribe your voice, then use AI for perception technology and interplay.
One of many challenges for these corporations is capturing customers’ voices in any type of surroundings, akin to loud cafes or places of work.
To deal with this, Refined Computing developed an end-to-end voice isolation mannequin that may perceive what you’re saying even in noisy environments. Chen mentioned that there are quite a lot of corporations engaged on voice understanding. He famous that at instances, gadget producers ship the voice to the cloud to get a clear output, however that’s not environment friendly.
The startup trains particular fashions to swimsuit the acoustics of a specific gadget and adapt to the consumer’s voice as an alternative of coaching one mannequin that works throughout units.
“What we discovered is that once we protect the acoustic traits of a tool, we get an order of magnitude higher efficiency than generic options. This additionally means we may give personalised options to the consumer,” Chen mentioned.
The corporate was based by Tyler Chen, David Harrison, Savannah Cofer, and Jackie Yang, who met at Stanford. Chen, Harrison, and Yang have been pursuing their PhD whereas Cofer was doing an MBA. They got here collectively in Steve Clean’s Lean Launchpad course, the place they labored on various interfaces for computing and began constructing Refined Computing.
Techcrunch occasion
San Francisco
|
October 13-15, 2026
“As we’re interacting extra with AI, we’re shifting in direction of a future the place we discuss with our units,” Chen mentioned. “However the apparent query is how a lot our units perceive us, the customers, in all of the environments the place we work each day. Be it a brilliant loud espresso store or a shared workplace the place there are different folks round you, and also you could be speaking about one thing personal — voice doesn’t work that manner at present,” he added.
The startup mentioned it may well run the mannequin only for voice isolation on some units, which is only a few megabytes in dimension and has 100ms of latency. The corporate can even run a unique mannequin to transcribe the voice and provides textual content output for different units. Chen mentioned due to its isolation mannequin, the corporate’s transcription mannequin can perceive customers higher, and in flip, creates a extra correct transcript.
Refined Computing mentioned that Qualcomm has chosen the startup as a member of its voice and music extension program. Which means the startup’s tech could be suitable with Qualcomm’s chips and be obtainable on units produced by OEMs.
The corporate has raised $6 million in seed funding led by Entrada Ventures, with participation from Amplify Companions, Summary Ventures, and angel buyers, together with founders like Twitter’s Biz Stone, Pinterest’s Evan Sharp, and Perplexity’s Johnny Ho.
Karen Roter Davis, Managing Companion at Entrada Ventures and a former director of an early undertaking at X (Alphabet), famous that voice AI is a loud house, and although interactions by way of this medium are selecting up, the general voice expertise just isn’t nice. She thinks that the startup’s give attention to voice isolation brings a unique perspective to the market.
“When you can debate whether or not AI will enhance or lower that point spent on a day-to-day foundation, we are able to all agree that advances in compute energy and machine studying / AI present alternatives for voice interface breakthroughs – if performed proper,” Davis mentioned. “Refined Computing is assembly folks the place they’re with voice interfaces that maintain up in excessive noise and excessive quiet, offering a voice expertise that’s dependable, straightforward, and enjoyable. It’s a recreation changer,” she added.
The corporate mentioned it has additionally partnered with a client {hardware} model and an automotive model — with out naming them — to deploy its options. However Refined Computing doesn’t wish to be only a mannequin provider to different corporations.
The startup additionally mentioned it plans to announce a client product that spans each {hardware} and software program subsequent 12 months, with out providing particulars.
