Compare talat and Google AI Edge Foresight

Foresight costs nothing and asks for no account. In English the transcription is quick and accurate: the words appear in italics as they are spoken and firm up a moment later.

The main part of the window is a notes editor. You type shorthand during the call and “Live Enhance” rewrites it into fuller notes. The transcript is one of four tabs beside it, along with Chat, Live Q&A and Summary. If you already take your own notes in meetings and want them tidied, that layout will suit you. In our test it showed “Live Enhance ON” but did nothing with a typed line until we switched the note style away and back; what it then wrote was decent.

It also has a library, called Sources, for PDFs, notes, audio, folders and Google Drive. Google describes Live Q&A as detecting questions asked during a call and answering them from those documents. We haven’t tested it properly, so we won’t judge it, but talat has nothing like it.

Foresight’s first run is a single setup screen that you can’t get past until the downloads finish: about 800 MB of “expert models” for speech recognition, query detection and text embedding, then the language model, which is 3.41 GB for Gemma 4 E4B. On a fast connection it took us about eight and a half minutes from opening the app to being able to use it. The speech models had arrived after two. You can’t decline the language model; the only choice is which size. The app itself is 376 MB, and it was using 6.9 GB of disk once setup was done.

We wiped a talat install and ran it from scratch on the same Mac and connection. talat fetches its speech models before your first recording too (1.4 GB of them, more than Foresight’s), but that is all it waits for: we were in the app after a minute and recording after three minutes forty, with the summariser still downloading in the background. Onboarding asks whether you want a summariser at all, and “Not automatically” downloads nothing extra.

The language model is Gemma 4. On our 32 GB Mac the choice was Gemma 4 E4B, the default, or Gemma 4 12B. There are no other models and you can’t bring your own. talat ships a local model for summaries and also lets you point it at Ollama, at a cloud provider with your own API key, or at your Claude Code or Codex subscription. You can change that choice at any time.

Foresight’s summary of our test recording, which was one person alone with a video interview playing, described a meeting in which “the team discussed how to push the boundaries of their product”. That recording was a jumble and we will test a real meeting. Small local models are limited, and the summariser talat ships doesn’t always write a great summary either. The difference is that talat writes from a transcript that knows who said each line, and lets you switch summaries off for a meeting, pick a bigger model, or summarise the same meeting again with a different one.

Each Foresight meeting has a chat beside it. Nine minutes into a recording we asked it “summary so far?” three times, and each time the answer was “I’m sorry, I don’t have enough details in the retrieved notes or transcript”.

A few more details come from the app’s own files, not from anything Google has published. It needs an Apple silicon Mac on macOS 14.2 or later. Gemma 4 comes in three sizes, E2B, E4B and 12B, offered according to how much memory the Mac has. The audio isn’t kept, only the transcript, and export is Markdown. talat keeps the recording on your machine and plays it back in sync with the transcript.