Google launches offline Mac transcription app powered by on-device model

Google has released an experimental note-taking application called Google AI Edge Foresight that transcribes meetings and audio files entirely on the device, with nothing sent to the cloud. The app is free, runs only on macOS, and is built on EmbeddingGemma 2, a model Google developed for local execution. For now it supports only Apple Silicon Macs, leaving Intel-based machines unsupported.

The model behind the scenes

EmbeddingGemma 2 is a compact embedding model designed to run on consumer hardware. Unlike large language models that demand heavy GPUs or a network connection, all audio processing, summarization, and retrieval happen in system memory. The practical upshot: no network latency, no service-availability dependency, and no API invoice at month end.

What the app does

Like competitors such as Granola and Wispr Flow, Foresight offers real-time transcription alongside a space for manual notes. Users can jot brief points during a meeting, and the app expands them into "polished notes" grounded in the full transcript. A built-in assistant can answer questions about what was said and can reference local files — PDFs, documents, plain text — attached as additional knowledge sources.

Privacy as default, not option

Google states that files, audio, and notes "never leave your computer." That is a strong claim in a market where most tools upload recordings to the cloud for processing, even when they offer a partial "private" mode. The on-device architecture makes the promise technical rather than merely contractual: there is no server to leak, no encryption key to trust.

Position in a crowded field

The space is dense. Granola has built a loyal following around a minimalist interface and smart summaries; Wispr Flow pushes fast transcription to mobile; every major productivity suite is adding a similar feature. Foresight's edge is architectural — fully on-device, no account, no subscription, no Google Cloud dependency. The drawback: availability limited to Apple Silicon Macs, and an "experimental" label warning that things may break.

What remains unknown

Google has published no performance metrics — no word error rate for transcription, no average latency, no memory or battery consumption figures. It is also unclear whether Windows, Linux, or mobile versions are planned, or whether model updates will arrive through the App Store or a separate channel. For now this is a niche tool for early adopters with a new Mac, not a mature replacement for daily meetings.