Prepare before you go offline
Offline transcription is simple once every required file is already on the Mac. Install LokalBot, open its model settings, and download at least one speech-recognition model. Process a short test recording while connected so macOS permissions, the model, and the output folder are all confirmed before the real meeting.
The app itself does not require an account or activation server. Once the selected files are present, recording, transcription, local summarization, search, and playback can run without internet access.
Choose the speech engine for the job
There is no universal best local model. LokalBot exposes several engines because language coverage, speed, timestamps, and hard-audio accuracy pull in different directions.
- IBM Granite Speech 4.1 2B is the recommended high-accuracy default.
- Parakeet TDT 0.6B v3 is the fast multilingual option, covering 25 languages and measuring around 190× realtime in the project's benchmark.
- Parakeet v2 focuses on English and can be useful when recall matters more than broad language coverage.
- Qwen3-ASR 1.7B covers 52 languages and dialects and is the heavier Qwen tier for difficult recordings; its download is about 3.2 GB.
- Qwen3-ASR 0.6B provides a compact global-coverage option at roughly 0.7 GB.
- Whisper large-v3 turbo covers 99 languages, supports word timestamps, and is a useful wide-language fallback at roughly 1.6 GB.
What happens during an offline meeting
LokalBot writes your microphone and the meeting application's output to synchronized local tracks. When the meeting ends, the chosen speech engine turns those files into text on the Mac. The built-in language-model backend can then write a recap without a server request. The transcript, recap, audio references, and search index remain in the local library.
Offline mode does not weaken recording-consent requirements. The app starts with manual recording as the default, and optional detection modes still leave the decision and legal responsibility with you.
What still uses the network
Model files and application updates have to come from somewhere, so those downloads require a connection. Optional Agent Mode setup also downloads its pinned runtime. Finally, if you replace the built-in inference backend with a non-loopback OpenAI-compatible URL, that configured service naturally needs the network and receives the request context. Loopback services such as a local Ollama instance can remain entirely on the Mac.
A useful offline checklist
- Download the speech model and one summarization model.
- Record and process a 30-second test with the meeting app you will use.
- Confirm both “Me” and “Them” tracks contain audio.
- Disconnect Wi-Fi and repeat the test.
- Keep enough free disk space for the models and meeting audio.