New startup builds dedicated microphone to replace keyboard typing with AI voice-to-text

A London-based startup called Relay is taking aim at the keyboard itself. Rather than relying on software alone, the company is building dedicated voice-to-text hardware paired with AI that can transform spoken words into written communication, letting people work without ever touching a keyboard.

The effort is being led by Cookie Xu and Raymond Zhu, both veterans of Nothing, the consumer electronics company. Xu previously headed AI and services there; Zhu worked as a hardware product manager. The pair's inspiration came from their own frustrations: they'd grown accustomed to using transcription software like Wispr Flow to capture work conversations, but office environments presented a constant challenge. Speaking clearly enough for software to understand while keeping the volume low enough not to disturb colleagues proved nearly impossible. "In the office, we never really could get the right volume for input," Zhu explains, noting he'd sometimes resort to holding an earbud in front of his mouth to lower the sound reaching nearby coworkers.

How Relay works today#

The company is releasing a macOS application immediately, powered by Google's Gemini models. Users assign a hotkey—the default is the Fn key—that activates voice dictation in any text field. Beyond simple transcription, Relay can execute contextual actions it calls "Skills": a user might ask it to draft a message to a colleague in Slack, and the software launches the app, identifies the recipient, and composes the text. Similar shortcuts work with Gmail and Google Calendar, and users can create custom integrations for other applications.

These capabilities come with privacy considerations. Because Relay needs to understand what's happening on screen to execute these actions, it requests permission to capture snapshots of your display. The company states that data is processed locally on the device and passed through Relay's servers only in transit to Google's infrastructure, with no retention by Relay itself—though granting an AI system visibility into your operating system introduces inherent security trade-offs similar to those raised by Microsoft's controversial Recall feature.

Current performance and limitations#

In testing, Relay's transcription accuracy fell short of what Google's newer Rambler feature delivers on Pixel phones. The software sometimes misses words and occasionally fumbles punctuation. It handles misspoken corrections reasonably well, recognizing when a user restates a sentence and discarding the false start. Emoji insertion proved inconsistent—flawless when requested directly, but prone to literal transcription when appended to sentences. One bright spot: the software generally distinguishes between messages and separate commands, such as when a user asked it to send a restaurant recommendation via Relay while requesting a Maps link in the same breath.