Fully offline AI meeting transcription with local Whisper on Mac, Windows and Linux

Fully offline AI meeting transcription means the speech model runs on your own computer, so the meeting audio is never uploaded to be turned into text. On a Mac or a Windows computer, note taker and transcription apps built on local Whisper (MacWhisper, Buzz, Vibe), open-source meeting note takers such as Meetily, and do-it-yourself setups with whisper.cpp all transcribe on the device. The summary is the second half: it stays offline only when a local language model writes it, for example one run through Ollama. Otherwise the transcript goes to a cloud AI.

Barside, the free desktop app we make, does the first half on your computer and not the second: it transcribes calls with Whisper on your computer, and its summaries use the AI account you connect. Everything below about other projects comes from their own README files and product pages as we read them on October 16, 2026. We ran no hands-on tests and took no screenshots, so treat each line as the project's own statement and open the linked page before you decide.

What does fully offline mean for a meeting transcriber?

A fully offline AI meeting transcriber does both of its jobs locally, on hardware you control. First a speech model turns the audio into text. Then a language model reads that text and writes the summary, the decisions and the action items. Several of the apps below can run the first job on your computer and send the second to a cloud AI, so "offline transcription" on a product page describes step one only.

Two things need the internet whatever you pick. An online call travels over Zoom, Teams or Meet, and each model has to be downloaded once. So in practice "fully offline" means that nothing the app does with your audio or your transcript needs a server. Before you trust that label, also check the extras: sync, cloud backup and what the app reports about its own use.

Which fully offline AI meeting transcription apps use local Whisper on Mac and Windows?

The fully offline AI meeting transcription apps with local Whisper for Mac and Windows that we found documented fall into three kinds: transcription apps built on Whisper, open-source meeting note takers that pair local speech to text with a local language model, and Barside, which transcribes calls locally and summarizes with a cloud AI you choose. The table sums up what each project's own page says, as of October 2026.

Whisper apps for Mac and Windows

MacWhisper. MacWhisper's site says it "transcribes audio, video, meetings, and system audio locally on your Mac using Whisper, Parakeet" and other models, and lists Offline among its features. Its feature list also names automatic recording of meetings in Zoom, Teams, Webex, Skype, Chime, Discord and more, automatic detection of when a meeting starts and ends, and Automatic Speaker Recognition, and the page we read does not mark which of these need Pro. For summaries and chat, the page names cloud providers you connect with your own API keys, such as OpenAI, Anthropic and Google Gemini, and also says: "Use local AI models with Ollama and LM Studio." With a local model chosen, both steps run on your Mac by the page's own description. As of October 2026 the page lists a free version and MacWhisper Pro at €64 per license, paid once, with lifetime updates. It is a Mac app, and the page mentions no Windows version.

Buzz. Buzz's README on GitHub opens with "Buzz transcribes and translates audio offline on your personal computer. Powered by OpenAI's Whisper." It runs on Windows, on a Mac with Apple silicon (the README says the last version for Intel Macs is 1.4.5), and on Linux as a Flatpak, a Snap or an AppImage. It lists live transcription from the microphone, speaker identification in transcribed media, exports to TXT, SRT and VTT, and a plugin system with plugins such as AI summary generation. We found no line in the README about recording the computer's sound, which is where the other people on a call arrive, so check that before you plan to record calls with it. The README also warns that the Windows app is not signed, so you get a warning when you install it.

Vibe. Vibe's README promises "Ultimate privacy: fully offline transcription, no data ever leaves your device" on macOS, Windows and Linux. It supports Whisper, Nemotron 3.5 and Parakeet TDT v3 models, and credits whisper.cpp and OpenAI's Whisper. It transcribes from the microphone and from system audio, lists speaker diarization, batch-transcribes files and exports SRT, VTT, TXT, HTML, PDF, JSON and DOCX. Summaries come two ways: through the Claude API, which sends the text to Anthropic, or with Ollama for local analysis and batch summaries. We found no separate meeting mode in its README.

Open-source meeting note takers that pair Whisper with a local model

A fully offline AI meeting note taker with local Whisper transcription also needs a local model for the summary, and Meetily describes that pairing. Meetily's README says "100% local processing. no cloud required." and "No data ever leaves your computer." It captures the microphone and system audio at the same time, transcribes meetings entirely on your device with Whisper or Parakeet models, and for summaries says "Ollama (local) is recommended, with support for Claude, Groq, OpenRouter, and OpenAI." The README offers a Windows installer, a macOS .dmg built for aarch64, and guides to build it from source on Linux. On Windows, "It requires an AVX2-capable x64 CPU; AVX-512 is not required." The Community Edition is MIT-licensed, and a paid Meetily PRO is offered to people who want enhanced accuracy and advanced exports.

Anarlog, which started as Hyprnote, calls itself "The privacy-first AI meeting notepad." Its README on GitHub says "On supported Macs, available built-in transcription models run on-device." and "Local Intelligence providers keep summaries and chat on your computer too." It says it listens to your device audio with no bot in the call, names local servers such as Ollama, LM Studio or Unsloth, keeps sessions, notes and transcripts in local SQLite, and says the cloud is opt-in and not required. Downloads are listed for macOS, Windows and Linux, but the README does not say where transcription runs on Windows and Linux, so we could not confirm that it stays on the device there. The community app is MIT-licensed, and the README adds that the team is now building another app, char.

How to build a DIY meeting transcriber with open-source Whisper and a local LLM

A DIY meeting transcription setup with open-source Whisper and a local LLM has four parts: a recorder, a speech engine, an optional speaker step and a model server for the summary. Each project below publishes its code under an open-source license, and each one asks you to manage models, versions and audio routing yourself.

1. Record the call. You need an audio file that holds your microphone and the computer's sound. How you capture the computer's sound differs between Windows, Mac and Linux, so plan this step first.
2. Turn speech into text. OpenAI's Whisper repository is the reference version, and it says: "Whisper's code and model weights are released under the MIT License." Its README expects Python 3.8 to 3.11 and the ffmpeg command-line tool. whisper.cpp, a "Port of OpenAI's Whisper model in C/C++", lists support for CPU-only inference, runs fully on the GPU through Metal on Apple silicon, and uses CUDA on NVIDIA cards and Vulkan as a cross-vendor way to use your GPU. Its platform list includes Mac, Windows and Linux.
3. Label the speakers, if you need names. OpenAI's README does not mention speaker labels. WhisperX adds speaker diarization from pyannote-audio, and asks for a Hugging Face access token and that you accept the user agreement of its diarization model. Its README is frank about it: "Diarization is far from perfect." whisper.cpp has an experimental speaker segmentation option called tinydiarize.
4. Summarize with a local model. Ollama installs on macOS, Windows and Linux and has a REST API at http://localhost:11434, so a short script can send it the transcript and a prompt. Its README mentions running models locally or in the cloud, so check that the model you pick runs on your own machine before you call the setup offline.

A prompt for step 4 can be short. This one is an example to adapt:

To test that the chain is offline, download every model first, disconnect from the network, and run a short recording through all four steps. If any step fails without a connection, that step was not local.

Hardware and accuracy trade-offs, as the projects state them

Larger Whisper models need more memory, and in OpenAI's table they run slower, with turbo as the exception at about eight times the speed of large. We found no single accuracy figure for meetings in either README. OpenAI's README says "Whisper's performance varies widely depending on the language." and publishes word and character error rates by language, measured on Common Voice 15 and Fleurs. The table below combines the numbers each README prints.

OpenAI says it measured those speeds by transcribing English speech on an A100, so speeds on your own computer can differ. A few more lines from the same pages help with the choice:

• English-only models. OpenAI offers .en versions of tiny, base, small and medium. It describes turbo as an optimized version of large-v3 with faster transcription, and says turbo is not trained for translation.
• No graphics card. whisper.cpp supports CPU-only inference, and its table lists about 388 MB of memory for base, so a laptop without a separate GPU is not ruled out. The README does not promise a speed for your machine.
• Bigger GPUs, faster batches. WhisperX says its faster-whisper backend needs less than 8 GB of GPU memory for large-v2 with a beam size of 5, and that batched inference reaches 70x realtime with large-v2, by its own measure, without naming the hardware. It also gives a command to run on the CPU, and on a Mac, with --device cpu --compute_type int8.
• Crosstalk. The WhisperX README warns: "Overlapping speech is not handled particularly well by whisper nor whisperx".
• Live text. whisper.cpp's stream tool samples the audio every half second and transcribes continuously, and the project calls its microphone example a naive example of real-time inference.

How Barside transcribes calls on your computer

Barside is a free desktop app for Windows, Linux and Mac that transcribes calls with Whisper on your computer. It records your microphone and the sound your computer plays, so it works with Zoom, Teams, Google Meet or any app that plays the call through your speakers or headphones, and nothing joins the meeting. Every 30 seconds it turns the new audio into text and adds lines to a note, each with its time and a You or Others label.

1. Install speech once. Open the assistant with Ctrl+Shift+A, select Assistant settings, and choose Install speech. It downloads Whisper and its English model, about 150 MB, and needs no AI key.
2. Add languages if you need them. Under Speech language, choose Install more languages · 141 MB for 34 more languages or Detect automatically, then Save assistant settings.
3. Open the call setup. Choose Transcribe a call in the assistant, click the microphone on the desktop bar, or press Alt+Shift+Space and type call.
4. Tell the people on the call. The card reminds you: "Let call participants know before recording."
5. Press Start recording. A banner reads Recording call · mic + computer, and the assistant shows Transcribing locally… while each piece turns into text.
6. Press Stop & save when the call ends. Barside transcribes the last piece, and the Call transcript note is done.

After the speech download, transcription needs no internet connection. Each piece with sound is also saved in the note as a Call audio clip, so you can check a line against the recording. Barside captures the computer's sound through the system's audio loopback on Windows, through PulseAudio or PipeWire on Linux, and on a Mac with macOS 14.2 or later, where macOS asks once to allow Barside under Screen & System Audio Recording. The full guide is transcribe calls on your computer.

What is not offline in Barside

Barside brings no AI model of its own, and its guides describe no local model option. Summaries, decisions and action items come from Meeting follow-through, which sends the transcript text to the AI you connect in Settings → AI: Claude, ChatGPT, Gemini or OpenRouter. With My plan, Claude Code or Codex runs on your computer signed in with your Claude or ChatGPT plan, and the model that reads the text is still that company's. The audio stays on your computer in every case. See meeting follow-up and action items for what the analysis produces.

If your summaries must stay local as well, three routes are open:

• Skip the AI step. With no AI connected, the note still holds the whole transcript and its audio, and the search bar finds it by any word.
• Use your own local model. The transcript is ordinary note text, so you can copy it into a local model you run yourself, such as one served by Ollama.
• Keep some work away from AI. In Library → Sources you can leave any workspace or note out of Library AI, and Barside cannot analyze a note that is left out.

Other limits to know: Barside labels lines by audio channel, so everyone except you is Others. It transcribes calls live and cannot transcribe an audio file you already have, which Buzz, MacWhisper and Vibe can. It uses Whisper's compact base models (base.en for English), which can mishear names, numbers and crosstalk, and it has no phone app.

Examples: which setup fits which person

These are made-up examples that show how the trade-offs play out.

• Example: Lena, a user researcher on a Mac, interviews customers under a non-disclosure agreement and wants nothing to leave her laptop. MacWhisper with a local model through Ollama or LM Studio keeps both steps on her Mac, by its page's description.
• Example: Marco, an IT lead on Linux, wants every part on hardware he controls and is happy in a terminal. Meetily built from source with Ollama, or his own chain of whisper.cpp and Ollama, matches what those READMEs describe.
• Example: Aisha, a consultant on Windows, keeps each client's notes, tasks and calls in one place. Barside transcribes her client calls on her computer into each client's workspace. For clients who accept a cloud AI she runs Meeting follow-through, and for the rest she leaves their workspaces out in Library → Sources and writes the recap herself.

A privacy checklist before you call a setup offline

1. Download the models first. A local model has to be on your computer before you go offline. Barside's speech engine comes from GitHub and its model from Hugging Face, once.
2. Find out who writes the summary. A local model server keeps it on your computer, and a named cloud AI receives the transcript.
3. Check sync and backup. In Barside, the optional Barside Plus cloud backup encrypts your library on your computer before upload, and cloud sync, if you turn it on, copies notes to your other computers through Barside's server.
4. Know what else the app sends. Barside counts which features are used: it sends barside.space the names of the features used that day, with the system and the Barside version, never notes, files or who you are. It is on by default, and a switch in Settings turns it off. The privacy page explains it.
5. Test with the network off. Run a short recording with no connection and see which steps still work.
6. Tell people you are recording. Recording rules differ between countries and states, and some require everyone's consent.

Barside's privacy guide to recordings and transcripts and what goes to the AI you connect cover the details.

How it compares

If both steps must stay on your computer, the projects above that document a local summary model fit better than Barside: MacWhisper on a Mac, Vibe and Meetily on Windows and macOS, or your own whisper.cpp and Ollama chain. Barside fits when you want offline transcripts of live calls inside your notes, next to tasks and reminders, and you accept a cloud AI for the summary or write it yourself.

• For recorded files and cloud services, read the best transcription apps.
• For meeting notes with no bot in the call, see AI note takers without a bot, the Granola alternative and the Otter.ai alternative.
• On Linux, compare AI meeting notetakers for Linux.
• For what an AI company receives from your notes, read AI note apps and client confidentiality.

Works well with

Once a transcript is in your library, these guides show what to do with it next.