Google ships offline dictation app
Google quietly launched an offline-first dictation app for iOS called Google AI Edge Eloquent that runs Gemma-based speech models on-device and edits speech into cleaner prose by removing filler and false starts. Because transcription and editing happen locally after model download, the app improves privacy and latency—features that matter for child-facing products where connectivity and data protection matter. But the app’s editing focus highlights a trade-off: systems that ‘fix’ speech for readability are not the same as diagnostic tools that must preserve raw utterances for instructional measurement. (techcrunch.com) (android.gadgethacks.com)
Most phone dictation sends your voice to a server, waits for text to come back, and then pastes the result. Google just shipped an iPhone app that does the speech work on the phone itself after you download the model. (techcrunch.com) (developers.googleblog.com) The app is called Google AI Edge Eloquent, and TechCrunch reported on April 7, 2026 that Google released it quietly on iOS. It is free, and it joins Google’s wider AI Edge push to run small models locally on phones instead of in distant data centers. (techcrunch.com) (developers.googleblog.com) The basic trick is simple: a speech model turns your voice into text, like a court stenographer typing as you talk. In Eloquent, that speech model is based on Gemma, Google’s family of lightweight models built to run on-device. (techcrunch.com) (developers.googleblog.com) Eloquent does not stop at raw transcription. When you pause, it rewrites the text into cleaner prose by cutting filler like “um” and “ah” and removing false starts, so the output reads more like a finished message than a verbatim transcript. (techcrunch.com) (android.gadgethacks.com) Gadget Hacks reports that the app includes four rewrite modes called Key points, Formal, Short, and Long. That means one recording can be turned into meeting notes, a polished email, or a compressed summary without speaking it again. (android.gadgethacks.com) Running locally changes two things users notice right away: speed and privacy. If the model is already on the phone, there is no round trip to a cloud server for the main transcription pass, and your spoken draft does not have to leave the device just to become text. (techcrunch.com) (github.com) That local-first design also makes the app more useful in weak-signal places like airplanes, school buildings, and spotty commuter routes. Offline software is less glamorous than a giant cloud model, but it keeps working when the network bar drops to zero. (techcrunch.com) (developers.google.com) Google has been building toward this for a while. In May 2025, TechCrunch reported that Google released AI Edge Gallery, another app for downloading and running models locally on phones, and Google’s developer posts have since expanded that edge stack across Android, iOS, and the web. (techcrunch.com) (developers.googleblog.com) The catch is that “cleaner” text is not the same thing as “truer” text. If a teacher, therapist, or researcher needs the exact pauses, repetitions, and abandoned phrases someone actually said, an app designed to smooth speech is editing away the evidence they would want to measure. (android.gadgethacks.com) (techcrunch.com) So Eloquent looks less like a neutral recorder and more like a pocket copy editor for speech. For everyday dictation, that is useful; for assessment, diagnosis, or any job that depends on the raw utterance, it is the wrong instrument by design. (techcrunch.com) (android.gadgethacks.com)
