Google’s offline dictation app
Google published an offline AI dictation app that lets users transcribe speech without live cloud calls — a big privacy and latency win for on‑device use cases. That move dovetails with broader industry work to shift certain AI functions out of the cloud and onto phones or local CPUs. (x.com)
Most phone dictation still works like a walkie-talkie: you speak, your audio goes to a remote server, and text comes back a moment later. Google’s new iPhone app flips that by downloading the speech model first and then doing the recognition on the phone itself. (techcrunch.com) The app is called Google AI Edge Eloquent, and Google’s developer site says it is built for “professional, ready-to-use text” from natural speech. TechCrunch reported on April 7, 2026 that it launched on iPhone as a free download. (ai.google.dev) (techcrunch.com) Speech recognition is the part that turns sound waves into words, like a court stenographer listening in real time. In Eloquent, Google says the app uses Gemma-based automatic speech recognition models that run after you download them to the device. (techcrunch.com) (developers.googleblog.com) That changes two old annoyances at once. If the model is sitting on the phone, you do not need a live internet connection, and you do not wait for a round trip to a data center before words appear. (blog.google) (support.apple.com) Google added a second layer after the raw transcript. TechCrunch says that when you hit pause, the app removes filler words like “um” and “ah,” cleans up self-corrections, and offers rewrite buttons such as “Formal,” “Short,” and “Long.” (techcrunch.com) There is also a split between local mode and cloud mode. TechCrunch reported that local-only processing can be used entirely on device, while cloud mode sends the cleanup step to Google’s Gemini models instead. (techcrunch.com) Google has been moving pieces of artificial intelligence onto local hardware for years. Pixel 4’s Recorder app in 2019 could record, transcribe, and search audio on the phone, and Google AI Edge now offers tools for developers building text, image, and audio features that run on mobile devices. (blog.google) (ai.google.dev) That push got bigger in 2025 and 2026 as Google expanded its small Gemma models for phones, tablets, and other devices. Google’s developer blog said Gemma 3n added on-device audio input in early preview in 2025, and Gemma 4 was pitched on April 2, 2026 for offline code generation, audio-visual processing, and mobile agents running on local hardware. (developers.googleblog.com 1) (developers.googleblog.com 2) Google is not inventing offline dictation from scratch. Apple’s own support page says iPhone dictation requests are processed on the device in many languages with no internet connection required, which means Google is entering a category users already know instead of creating a new one. (support.apple.com) What Google is testing here is a different product shape: not just “type with your voice,” but “speak messily and get cleaned-up prose back.” If that works well enough, the same trick fits email, meeting notes, field reports, and any job where a person can talk faster than they can type. (ai.google.dev) (techcrunch.com)
