Questions people ask before installing

Yes, completely, once the AI models are downloaded. Airplane mode is fine.

Yes. Hark is MIT-licensed open source. No tiers, no seats, no minute limits, no upsell.

The GPU tier uses Whisper large-v3-turbo, which is state of the art for English and most major languages. CPU tiers use smaller Whisper models and are still good; the AI cleanup pass repairs many of their mistakes.

No. Hark records what your microphone hears and what the meeting app plays. Nobody in the call can tell.

Not on Windows: you pick the meeting window before recording and Hark captures only that application's audio (plus your mic). Choose 'Everything playing' if you want all system audio. On macOS all system audio is recorded.

Yes. Open the recording from the Library while it runs: captions appear as people talk and running notes - topic, decisions, action items so far - update every minute or two. The full transcript with speaker names arrives when you stop.

With the CUDA build about 10 minutes for an 88-minute call, everything included. On a CPU-only machine Hark measures each speech model's real speed and picks the best one that finishes within the recording's own length; speaker detection runs across all cores in parallel with transcription.

Yes. Library > Import takes audio or video files (a phone voice memo, an old Zoom .mp4) and runs the same pipeline: transcript, speakers, summary, search.

Zoom, Microsoft Teams, Google Meet, Webex, Discord, Slack huddles, FaceTime and GoToMeeting by window. Everything else by audio activity (mic and speakers both live). You can add detection patterns in Settings, and the hotkey works anywhere.

Yes. Start a recording manually; speaker identification works from the microphone alone.

pyannote segmentation splits the audio into speaker turns and TitaNet computes a voiceprint per speaker, all on your machine. Your own voice is recognised because it arrives through the microphone track. Name someone once and Hark will suggest their name next time; it never applies a name without you confirming.

Any 64-bit Windows 10/11 or macOS 13+ computer with 8 GB RAM. An NVIDIA GPU with 6 GB+ VRAM, or Apple Silicon, unlocks the larger models and makes everything several times faster.

Models: 1 to 7 GB depending on tier. Recordings: about 35 MB per hour of audio, plus roughly 300 to 600 MB per hour if you record screen video.

Yes. Point Hark at Ollama, LM Studio or any OpenAI-compatible server in Settings. Transcription still happens in Hark.

Export everything as a .hark bundle and import it on the other machine, or move the whole data folder.

Code-signing certificates cost money the project doesn't have yet. Check the SHA-256 on the release page, or build from source.