Google quietly launches AI Edge Eloquent, a free offline-first AI dictation app for iOS powered by on-device Gemma models

Google on April 7 quietly released an experimental iOS app called Google AI Edge Eloquent — a free, offline-first dictation tool that uses on-device Gemma-based speech recognition models to transcribe speech, strip filler words, and polish text without sending audio to the cloud. The app, which does not require any subscription, competes in an increasingly crowded category that includes Wispr Flow, SuperWhisper, and Willow.

Eloquent's design reflects a growing emphasis on privacy-preserving, on-device AI. Once a user downloads the Gemma-based automatic speech recognition model, the app can function entirely locally, with no data leaving the device.

The app displays a live transcription as the user speaks, and when they pause, it automatically removes filler words like "um" and "ah" and restructures the transcript into cleaner prose. Users can also choose from formatting options including "Key points," "Formal," "Short," and "Long" to shape the output.

For those who prefer enhanced quality, Eloquent offers a cloud mode that routes text cleanup through Google's Gemini models. The app can optionally import keywords, names, and jargon from a user's Gmail account to improve transcription accuracy for specialized vocabulary.

Google's App Store listing initially referenced an Android version and keyboard integration, but those references were removed in an update on the same day, suggesting the Android rollout is still pending.

The launch is framed as an experiment by Google, but it signals the company's interest in applying its Gemma on-device model work to consumer productivity tools. If the app gains traction, it could accelerate Google's rollout of improved transcription features across its broader Android ecosystem, where on-device speech recognition would be particularly valuable in offline or bandwidth-constrained environments.

Read the original reporting at TechCrunch.