Eloquent

Most people speak faster than they type when they don't have to plan every sentence in advance, yet traditional dictation tools transcribe every pause, filler word, and mid-sentence correction verbatim. Google AI Edge Eloquent solves this problem for verbal thinkers by turning unfiltered spoken thoughts into clean, structured text directly on a user's device. Unlike cloud-dependent transcription apps, Eloquent runs speech recognition and language models locally without requiring an internet connection, keeping voice notes fast, private, and available offline.

Built by Google AI Edge Labs, Eloquent and its sister app Foresight serve as flagship reference apps demonstrating what developers can build with Google's core on-device AI SDKs. After launching Eloquent on the App Store, real-world adoption took off: the app climbed into the Top 20 on the App Store, earned over 100 press mentions, and has saved users millions of minutes of typing. Usage patterns also surprised the team, as users began recording 30-to-60-minute sessions rather than short voice notes, prompting an update that unlocked unlimited dictation duration. Soon after, the team was invited to showcase the Gemma 4 12B model in Eloquent on launch day. Because a 12-billion-parameter model is too large to run on a mobile phone, meeting the launch meant transforming their iOS mobile app into a production desktop application within a 10-day window.

Bridging native on-device AI pipelines with Flutter

When the team started building Eloquent, none of the engineers had prior experience with Flutter or Dart. Thanks to Dart's approachable syntax, Flutter's declarative widget model, and thorough documentation, the team became productive right away, using Flutter to orchestrate complex on-device AI pipelines alongside responsive, interactive interfaces.

At the core of Eloquent's local intelligence is LiteRT-LM, Google's on-device inference framework for running Gemma models efficiently on consumer hardware. Eloquent and Foresight's transcription, embedding, and retrieval pipelines run on the same code that powers Google's newly released MediaPipe Tasks, including Audio Transcriber, Universal Embedder, and SemanticRetriever. Rather than maintaining separate native UI and model-execution layers for each operating system, the team used Flutter and dart:ffi as a single bridge to the native C++ API. That architecture allowed one shared Dart library to power model management, audio transcription, and UI state across platforms, and any Flutter developer can build on top of the same underpinnings with the Flutter MediaPipe package, supporting all 22 MediaPipe tasks.

Shipping on desktop in 10 days

When the 10-day countdown to the Gemma 4 12B showcase began, the team's Flutter foundation paid off immediately. Building directly on their iOS codebase, designers and engineers adapted the interface for larger desktop screens, added window resizing and desktop keyboard shortcuts, and integrated local 12B model execution on macOS in time for launch day.

Flutter also made it straightforward to craft expressive, custom UI interactions that make on-device AI feel transparent. A signature example is Eloquent's diff animation, which visually highlights the differences between a user's raw spoken words and the polished version generated by the local model.

That modular Flutter architecture became the launchpad for the team's next application, Google AI Edge Foresight, an AI-enhanced meeting and note-taking companion powered by EmbeddingGemma 2 and Gemma 4. Foresight lets users jot down shorthand notes during a meeting and expands them into complete, structured notes alongside live transcripts and summaries.

Because Eloquent was built as reusable Dart and Flutter packages, Eloquent and Foresight share roughly 60% of their overall infrastructure stack. Foresight reused Eloquent's entire design library, including the diff animation, along with shared middle-tier packages for model downloading, local encryption, and LiteRT-LM orchestration, cutting an estimated 80% of initial engineering time compared to building that low-level C++ FFI, model-management, and UI infrastructure from scratch.

"Flutter is not only great for UI, you can use it as a single bridge to native pipelines as well. Don't limit yourself to thinking Flutter is a great UI tool for you to build cross-platform, but also think of ways to optimize your entire stack using Flutter."

- Google AI Edge Engineering Team

Key results and business impact

By selecting Flutter, the Google AI Edge team delivered the following results:

• Mobile to desktop in 10 days: Adapted the iOS application for macOS desktop layouts, window management, keyboard shortcuts, and local Gemma 4 12B inference within a 10-day launch window.
• 60% shared infrastructure stack: Reused Eloquent's Flutter design system, interactive UI components, and Dart-to-C++ middle tier as the foundation for Google AI Edge Foresight.
• Nearly 80% engineering time saved: Accelerated Foresight's development by sharing model-management, speech, and UI libraries rather than rebuilding low-level plumbing from scratch.
• Strong App Store adoption and rapid iteration: Reached the Top 20 on the App Store with 100+ press mentions and millions of minutes of typing saved, shipping unlimited dictation support within three weeks despite the team starting with zero prior Flutter experience.
• Single cross-platform bridge to C++ ML runtimes: Connected LiteRT-LM, MediaPipe, and local Gemma models through dart:ffi and a shared Dart library across mobile and desktop targets.

Ready to build on-device AI experiences in your own Flutter apps? Get started with the flutter_mediapipe package on pub.dev, explore the source code in the flutter-mediapipe GitHub repository, and consult the Google AI Edge and LiteRT-LM documentation.

To experience the Google AI Edge Labs showcase apps firsthand, visit Google AI Edge Eloquent, download Eloquent on the App Store, or explore Google AI Edge Foresight.