Google Quietly Launches Offline AI Dictation App to Rival Wispr Flow

Google slipped a fully functional AI dictation app onto the App Store without a press release, without marketing, and without any public announcement. Named Google AI Edge Eloquent, the app appeared in the iOS App Store as early as April 6 and launched with zero ratings, a precise gauge of how deliberately under-the-radar its debut was.

The app's defining architecture is offline-first. All speech recognition runs locally on the device using Gemma-based ASR models downloaded directly to the phone, meaning audio never reaches Google's servers unless the user chooses otherwise. That design makes Eloquent well-suited to the categories of work where uploading voice data to a cloud service is inadvisable: legal transcription, clinical notes, internal corporate communications, and any environment where connectivity is unreliable. An optional cloud mode routes text cleanup through Gemini's server-side models for potentially higher-quality polishing, but users can disable it entirely.

Eloquent transcribes speech in real time and, when the user pauses, strips filler words including "um" and "uh" along with mid-sentence self-corrections, then copies the cleaned text directly to the clipboard. Four transformation modes allow further reshaping: Key Points, Formal, Short, and Long. The app builds a personal vocabulary dictionary from the user's speech and can optionally pull frequently used words, names, and jargon from recent Gmail history. Each session logs words per minute and total words dictated.

The on-device accuracy trade-off is real. Eloquent relies on Gemma 4's mobile-optimized E2B (2 billion parameters) and E4B (4 billion parameters) variants, which Google says can run with under 1.5GB of memory on some devices. That efficiency is genuine, but it comes with ceiling limits compared to the server-side large language models used by cloud-first competitors.

Those competitors charge substantially more. Wispr Flow's Pro tier runs $12 per month (roughly $10 per month billed annually), with a free Basic plan capped at 2,000 words per week on desktop and 1,000 on iOS. SuperWhisper charges $249 for a lifetime license. Eloquent is entirely free with no usage caps, undercutting both rivals while also outperforming Apple's native iOS dictation, which provides no filler-word removal, no text transformation, and no vocabulary learning. A widely circulated Reddit discussion from February 2026, documenting what users called a "Wispr Flow Trust Gap" around post-payment quality degradation, had already begun eroding the category leader's standing before Eloquent arrived.

The choice to release Eloquent under the Google AI Edge brand, rather than through Google's consumer product channels, carries strategic weight. Google AI Edge is the company's designation for on-device AI, home to the Google AI Edge Gallery app that lets developers run Gemma models locally. That Gallery received an update alongside Gemma 4's early-April launch, adding "Agent Skills," described as the first on-device multi-step agentic workflows running entirely locally. Analysts at The Next Web noted that positioning Eloquent within this brand signals a dual purpose: a developer and enterprise showcase for on-device Gemma capabilities, and a direct consumer product simultaneously.

At launch, Eloquent is iOS only, requiring iOS 16.0 or later. An Android version is referenced in the App Store listing but has not appeared on Google Play. Whether the stealth release is a technical preview or the opening move in a larger consumer push remains Google's to clarify, on its own timeline.