Performance and your hardware
Daisy runs entirely on your machine. Recording, transcription, and speaker labels never leave your computer. That's the point of Daisy — and it also means your hardware sets the pace.
The expensive part happens while the meeting runs.
Transcription happens during the meeting
Daisy transcribes as you talk, and that text is your transcript. When you press ✦ Finish & summarize, it keeps what it already has. The transcription cost is spread thinly across the length of the call.
The default engine is chosen to make that cheap:
- Windows and Linux — a bundled Parakeet model that runs on your CPU, so it needs no graphics card and leaves the one you have to your meeting app.
- macOS 26+ — Apple's own on-device speech engine, built into macOS.
- Whisper — used when you choose it, or as the fallback. This is the one that uses your graphics card when it can, and runs more slowly on the CPU when it can't.
What still takes time when you stop
The main steps:
| Step | What it costs |
|---|---|
| Echo cancellation and audio cleanup | Scales with meeting length |
| Speaker labels | Runs on your machine |
| Filling coverage gaps | Runs only if the live pass missed a stretch |
| The AI summary | Often the longest step on a modest machine |
That last row is the one to pay attention to — see below.
When a full transcription pass does run
There are two cases where Daisy transcribes a whole file from scratch, and both show progress as transcribed n/m chunks:
- Imported audio — a voice memo or an existing recording has no live pass behind it. See Importing audio and notes.
- A session recorded with no model installed, or one where the live transcriber was paused mid-meeting because the machine couldn't keep up. The audio is intact, and Daisy does the work now.
This is the case where a GPU and a smaller model both help.
The summary is a separate step
Summaries, chapters, and analysis need an AI model, and on a modest machine that's the slowest part of the whole flow:
- An on-device AI model — Daisy's built-in one, or LM Studio or Ollama — runs at the speed of your hardware, and is often the slowest step on a laptop. The built-in model wants a graphics card with 6 GB of memory, or 16 GB of system memory.
- A cloud key (Anthropic, OpenAI, or Groq) is usually much faster. Only the transcript text is sent, never your audio, billed to your own key. See Configuring AI providers.
- No AI at all — use the copy-paste workflow and summarize in your own ChatGPT or Claude. See Running Daisy fully offline.
A cloud key speeds up summaries only, and never sees your audio.
If your machine feels slow
- Keep the built-in model. On Windows and Linux, the bundled Parakeet model is the fast path and needs no GPU. On a Mac, Apple's engine is. Switching to a large Whisper model trades a lot of speed for accuracy you may not need on clean meeting audio.
- If you do run Whisper, keep your graphics driver up to date. Without a working driver, Whisper falls back to the slower CPU path.
- Hide the captions if the UI feels heavy. Settings → Recordings → Live captions is display-only; turning it off stops the recording view repainting constantly. The finished transcript is the same.
- Tune the engine, if you know what you're changing. Settings → Providers → Advanced → Transcription tuning has settings for each engine, each group with a Reset to defaults button. The defaults suit most machines; this is there for the ones they don't.
- Bring a cloud key for fast summaries if the AI step is what feels slow.