Google Quietly Releases an Offline AI Dictation App for iOS Powered by Gemma

Google released a voice dictation app for iOS that operates entirely offline, using its Gemma AI models to convert speech to text without sending data to the cloud. The release was quiet — no major announcement, no blog post — which has become something of a pattern for Google's incremental Gemma deployments. The app competes directly with dedicated dictation tools like Wispr Flow that have built audiences among knowledge workers who dictate heavily.

What the App Does

The app's core function is standard: speak, and text appears. What differentiates it from conventional mobile dictation — including Apple's built-in system — is that the AI model running the transcription lives on the device. No audio is sent to Google's servers. No internet connection is needed. The model processes speech locally and returns text.

On-device dictation is not new in concept, but the quality has historically been limited by the computational constraints of consumer hardware. Running a capable AI model locally on an iPhone requires careful optimization — quantization, architecture choices, memory management — that earlier on-device models couldn't deliver at acceptable accuracy. Gemma's architecture, designed with on-device deployment as an explicit goal, changes that calculation.

The Wispr Flow Comparison

Wispr Flow has become the preferred dictation tool for a growing cohort of AI-forward knowledge workers who have moved from typing to voice for most of their text input. Its differentiators are speed, accuracy on technical vocabulary, and integration with macOS and iOS text fields across applications. Google's offline app enters this space with a significant brand advantage — and the trust infrastructure of being built by the company that has led speech recognition research for two decades.

Whether the accuracy matches Wispr Flow's current standard is the critical unanswered question. Google's initial release offers parity on the privacy dimension (both are effectively local or edge-based), but dictation software lives or dies on recognition accuracy on domain-specific vocabulary, accented speech, and mixed-language input. Those details will surface in extended use.

The Broader Gemma On-Device Strategy

The dictation app is the latest data point in Google's systematic push to deploy Gemma inference on consumer hardware. Gemma 4, released last week, brought frontier multimodal capabilities to on-device deployment. The dictation app brings speech recognition to the same stack. The pattern suggests Google is building toward a device-level AI platform — one where multiple AI functions (vision, speech, text) run on Gemma locally, reducing dependency on cloud connectivity and addressing the data privacy concerns that have made some users cautious about AI features in consumer software.

For Apple, whose own on-device AI ambitions have been more prominent in marketing than execution, a capable Google dictation app running locally on iOS is a pointed competitive signal.