Wispr Flow preps voice AI’s next breakthrough

Tanay Kothari set out to build a real-life version of Tony Stark's AI assistant, JARVIS. What he ended up creating instead is a better foundation for how humans will interact with AI.

Kothari founded Wispr in 2021 with a team of PhDs in neuroscience, machine learning and signal processing. The original product was a wearable that converted neural signals from silent speech into text, letting users control their devices without saying a word. The technology actually worked, but after connecting it to ChatGPT, Siri and Alexa, the team hit a wall. The thoughts in your head are too rambling to be useful as raw input, and no consumers could immediately see what to do with such a device.

"We were asking people to change two behaviors at once, wearing a new device and using voice as their primary input, and the market simply wasn't ready for either," Kothari told The Deep View.

In mid-2024, Kothari made the call to pivot. The company went from 40 people to five. Wispr had been quietly building a second product in parallel called Flow, a voice-to-text tool designed to work across any application on your computer. The breakthrough insight came not from the hardware product the team set to build, but from watching employees talking to their machines through $10 microphones. So that's what Kothari pivoted the company around.

"The best product decisions come from humility, not conviction," Kothari said. "Conviction almost killed our company."

Flow is now used by more than 270 Fortune 500 companies. At Clay, a sales intelligence software company, roughly half of their go-to-market and revenue organization adopted Flow within a single month with no top-down mandate, driving 52% faster customer response times and an estimated $3.08 million in annual cost savings.

The product works by layering AI reasoning on top of voice input. Unlike standard transcription tools that treat voice as a simple audio-in, text-out problem, Flow uses screen context, conversation history and user corrections to resolve the ambiguities that trip up conventional speech recognition, things like mumbled names, noisy rooms or rapid speech. Every time a user corrects an output, the system learns their vocabulary, preferences and intent.

Try Wispr Flow and see how voice can replace typing

"The goal is simple," Kothari said. "Correct the same mistake twice, and you never have to correct it again."

Some of the most compelling use cases come from users with accessibility challenges. Ron Kort, a civil rights attorney in New York born with Spinal Muscular Atrophy, now does all of his legal writing by voice and finishes work in a fraction of the time that typing required.

Looking ahead, Wispr is training its own large-scale voice models built for personalization and moving into what the company calls "voice-to-action," where speaking doesn't just produce text but triggers real workflows like drafting follow-ups, scheduling meetings or updating documents. Beyond that, the roadmap includes a developer ecosystem and wearable hardware.

But Kothari is deliberate about sequencing. "Each phase has to earn the next," he said. "Right now, we're heads down on making voice input so good that a billion people trust it enough to put down their keyboards."

Experience Wispr Flow’s voice-first workflow