Skip to content

Use OpenWhisper

Meeting mode

Capture a call, configure optional AI insights, and export the results.

On this page

Meeting Mode captures your microphone and the call's system audio, builds a transcript, and can add live AI notes and a final report. It works on Windows and macOS 13 or later; Linux system audio is a preview.

Prepare audio capture#

Choose your microphone and a speech engine that supports meetings. Meeting Mode uses the same microphone ranking as dictation (Settings → Dictation → Recording). On macOS, allow Microphone and Screen & System Audio Recording in Privacy & Security.

Before your first meeting, setup checks for missing speech runtimes and models and offers to download them; downloads need a click and respect your offline settings. Then choose Check microphone and system audio, speak, and play sound through the output your calls use. The test audio is discarded. You can start even if setup is incomplete, but missing models or audio may leave you without a usable transcript, and on macOS a denied system-audio permission offers a microphone-only start. Reopen the checks any time with Meeting setup….

During a meeting, the dashboard's More meeting options → Audio check tests the microphone while you speak and system audio while something plays, with live levels for each and a countdown. The check detects a signal but can't tell its source, so confirm both you and the other side are present. Capture-health and incomplete-audio warnings stay visible if anything goes wrong during a recording.

Choose a speech engine and speaker labels#

Meetings use your dictation engine and model, so there's one speech engine to set up. To give meetings their own, open Settings → Meeting Mode → Voice & speakers and set Speech engine to A different engine; Runs on and Meeting speech model appear then. Qwen3-ASR, Parakeet Unified EN 0.6B, and the API engine are for dictation and uploads only (Parakeet Unified's compact TDT-CTC 110M model does handle meetings): if dictation uses one of them, meetings use the default meeting model, and Settings and the Meeting tab say so. See Choosing a model.

If dictation uses a paired host, meetings do too; to use a host only for meetings, choose A different engine, then Runs on → Remote computer. Microphone and system audio go to the host for transcription; if the connection drops, capture continues locally and transcription retries. See Remote engines.

Speaker identification is separate from the speech engine:

  • Off labels lines as Me and Others by audio channel.
  • On-device (the default) uses WeSpeaker. Its roughly 27 MB model downloads from Hugging Face the first time it's needed; if it's unavailable, the meeting keeps Me/Others labels and the banner says why (a network error, Hugging Face set to Never connect, a refused download, or a model that won't load). Voice & speakers shows whether the model is ready, not downloaded yet, blocked, or failed, with Download now. Speakers are told apart on system audio only. The model gets ready in the background, so transcription starts at once; system-audio lines spoken before it's ready stay on Me/Others.
  • OpenAI uploads system audio after the meeting to relabel the saved transcript. It needs an OpenAI key and explicit consent, never uploads your microphone, and falls back to on-device labels once OpenAI retires its diarization model.

Add optional AI insights#

Settings → Meeting Mode → Intelligence chooses who writes live notes and the final report:

EngineSetup
Pi (default)Install from Models & storage → Downloads → Components, then pick its text model and key.
Claude Code, Codex, or OpenCodeSelect an agent you've already installed and signed in to. Its meeting passes count toward that agent's plan or API usage.
OpenCode SDKInstall from Components, then select it under Agent core.

The Insights model card holds the model and two Thinking levels: one for live passes and one for the final report (Low and Medium by default). Each lists only the levels that model accepts. Local servers such as Ollama, LM Studio, vLLM, llama.cpp, and LocalAI work here too, and are checked for tool support before a meeting relies on them.

If the chosen engine is unavailable, install or update it, pick another, or keep recording without AI insights. Recording never depends on AI.

To keep insights going when they fail, pick Claude Code, Codex, or OpenCode as the Backup agent on the same page. If AI insights stop working during a meeting, or can't start, it takes over, catches up on the meeting, and writes the final report. The dashboard says that it took over and why.

Use Fast judgments#

Settings → Meeting Mode → Fast judgments adds quick, focused checks: topic changes, spoken instructions, highlights, a questions radar, citation checks, meaning-based history search, and insight review. Under Judge, choose:

  • TypeSafe or OpenRouter, which run TypeSafe's Jev model with that service's key, or
  • Your text model, your AI insights or AI cleanup model. Test reports how long it takes.

Record and review#

The Meeting Mode tab has two sides. The left column holds Start Meeting, the speech engine in use, the AI insights switch with its model, and an optional Meeting brief. The brief guides AI notes and the live copilot, and you can edit it mid-meeting. The right side shows your three newest meetings to open; while a meeting runs, an on-air card there has the clock, Pause and End (red while capturing, amber while paused), with Open dashboard and Copy guest link in place of Start. When it ends, the wrap-up steps appear on the right. In a narrow window the column sits above.

The left column always sets up your next meeting. While the last one is still wrapping up, Start Meeting is dimmed with a line saying why, and unlocks as soon as the wrap-up can continue in the background; starting then sends it there. If a meeting's final insights failed, starting the next files it in Past Meetings, or choose Keep for later. Find results in Past Meetings, or on View → History Calendar alongside your dictations. Check names, decisions, and commitments against the transcript before sharing a report.

To fix a passage, select it in the dashboard's Conversation view and type a replacement. Choose Only this occurrence for a focused fix, or apply it across the meeting.

Export a meeting#

Export from the dashboard or Past Meetings. Markdown can include the transcript and AI insights; plain text is the transcript; JSON keeps the structured state. Exports reference recordings rather than including audio. See Export support, and use Settings → Models & storage → Backup & restore to keep records and recordings together.

Checked against OpenWhisper 2.8.97.App documentation
Still stuck?

Get help from the community, or tell us what went wrong.