Google Launches Offline AI Meeting Transcription App for macOS
Google has launched Google AI Edge Foresight, a free experimental macOS application that provides local, offline meeting transcription and summarization powered by Gemma models.
The market for automatic meeting transcription and summarization has surged over the past year, with productivity tools like Granola and Wispr Flow becoming indispensable for professionals. However, security concerns regarding sensitive conversations and cloud uploads have led Google to introduce Google AI Edge Foresight, a free experimental macOS app running entirely offline.
The Local AI Engine
Foresight relies on on-device processing to ensure complete data privacy. EmbeddingGemma 2, a lightweight 740-million-parameter multimodal model, maps text, images, video, and audio into a unified vector space. This enables real-time semantic retrieval, acting as a local RAG system with zero network latency.
For summarization and communication, users can download local Gemma 4 models ranging from 2B and 4B to the powerful 12B version. These models analyze transcripts, synthesize notes, and answer queries directly on the user's hardware.
Streamlined Workflow
Google adopted a split-screen interface popularized by competitors: one side shows the live transcript, while the other captures rough manual notes. After the meeting, the system cross-references these notes with the full transcript to generate polished summaries.
Users can also link local folders, calendars, and documents in PDF, Office, Google Docs, or Markdown formats. The built-in Ask tab allows users to query past discussions and connected files in real-time.
Limitations and Availability
Currently, Foresight is exclusively available for macOS and optimized for Apple Silicon processors (M1 and above), omitting Intel-based Macs. Installing the app alongside the 12B Gemma 4 model requires roughly 7 GB of local storage. Furthermore, the app currently supports live audio streaming from physical meetings or video platforms like Google Meet, Zoom, and Teams, but lacks support for pre-recorded audio files. While Gemma models support Hebrew, early testers report that live transcription performs best in English.