Skip to content

All pages

Turning meeting and lecture recordings into text on your own machine

5 tools, Last updated 2026-09-16

What separates these tools is not the accuracy claim on the landing page but three things: which speech model runs underneath, whether Korean is on that model's supported list, and which machines it runs on. Many are Apple Silicon only, and file transcription and live dictation are different features even inside the same app.

One thing trips up Korean users in particular. NVIDIA Parakeet, the default engine in several of these tools, covers 25 European languages and no Korean. Feed it a Korean meeting with the defaults and the transcript is garbage, which is easy to blame on the tool. Switching to a Whisper model or Qwen3-ASR fixes it.

How this was checked. Written on 2026-09-14 from each tool's GitHub README, official docs and the speech model cards. Entries we installed and fed a Korean recording are marked "Installed and tried"; the rest are "From documentation". Anything the docs do not state is written down as not stated.

At a glance

InstallMeetilyFrom documentationOpenWhisprFrom documentationQwen ScribeFrom documentationYapFrom documentationVoiceStudioFrom documentation
InstallInstaller (.dmg/.exe); Linux builds from sourceInstaller (.dmg/.exe/AppImage)make from source (needs Python, ffmpeg)One brew line or a .dmgInstaller (.dmg/MSI/AppImage) or Docker
KoreanUse Whisper; Parakeet has no KoreanPick Korean under the 59-language Whisper engineKorean is on Qwen3-ASR's official 30-language listFollows the system locale; Korean unconfirmedWhisper-family engines; Korean not named in README
Key constraintSpeaker labels, exports are PRO ($10/mo)No speaker ID on Intel MacsApple Silicon only; one file at a time, 4GB maxmacOS 26+, no file transcription8GB RAM minimum; Intel Macs need a remote backend

Meetily

Meeting transcription and summaries that never leave the machine. 30,894From documentation

Pick this when. When you want to drop in a whole meeting recording and get the summary in the same pass, and you are fine installing a local LLM (Ollama) for it.

Installation

  1. macOS: download the Apple Silicon .dmg from the releases page and drag it into Applications.
  2. Windows: download the x64 installer (.exe) from the releases page. The CPU must support AVX2.
  3. Linux: no installer, source build only. Rust and Node.js must be installed first.
  4. For local summaries, install Ollama separately; otherwise enter a Claude, OpenAI, Groq or OpenRouter key.

Korean-language usability

The transcription engine is a choice between Whisper (tiny through large-v3) and Parakeet. Parakeet is a 25-language European model with no Korean, so use Whisper for Korean meetings and large-v3 when accuracy matters. The official site mentions multi-language support with auto-detection but publishes no list of languages. We could not confirm a Korean UI.

Real constraints

  • Speaker identification, PDF/DOCX export and language auto-detect are PRO features ($10 per user per month, billed as $120 a year). The community edition is free under MIT.
  • Official minimum spec is 8GB RAM, 4 cores and 10GB storage; recommended is 16GB, 8 cores and 50GB. large-v3 is heavier than that.
  • Linux is still source build only.

Checked on 2026-09-14, Checked: From documentation. Written from the official docs and README. Not installed and tried by us.

OpenWhispr

Desktop dictation that types where your cursor is, using local or cloud speech models. 8,321From documentation

Pick this when. When you want to hit a hotkey mid-document, speak, and have the words land at the cursor. The same app also transcribes meeting files.

Installation

  1. macOS (Apple Silicon and Intel) and Windows: .dmg and .exe installers from the releases page.
  2. Linux: AppImage, .deb, .rpm and .tar.gz.
  3. Source builds need Node.js 24 or newer: git clone, then npm install and npm run dev.
  4. Local Whisper is GPU-accelerated on Metal, CUDA and Vulkan. Cloud models need your own API key for each service.

Korean-language usability

Language is set under Settings, Preferences, Language, with auto-detect as the default. Coverage differs per engine: Whisper 59 languages, Parakeet 26, AssemblyAI 7. Parakeet is a European-language model, so Korean means choosing Whisper. The docs themselves advise naming the language rather than relying on auto-detect for short or noisy clips.

Real constraints

  • Intel Macs lose live speaker identification and voice fingerprinting: both depend on ONNX Runtime, which stopped shipping macOS x86_64 builds at 1.24.
  • Notes require signing in (free). Cloud engines bill you at the provider for what you use.
  • MIT licensed, no usage tracking.

Checked on 2026-09-14, Checked: From documentation. Written from the official docs and README. Not installed and tried by us.

Qwen Scribe

Transcription and dictation that stay on an Apple Silicon Mac. 213From documentation

Pick this when. When you are on an Apple Silicon Mac and want long recordings transcribed with SRT timestamps in the same pass.

Installation

  1. Apple Silicon Macs only, macOS 14 or newer.
  2. Install Python 3.12 or newer, ffmpeg (brew install ffmpeg) and the Apple command line tools (xcode-select --install) first.
  3. Clone the repo, run make setup to create the virtual environment, then either make app to build the app or ./run.sh to serve it at http://127.0.0.1:8990.
  4. There is no .dmg installer. You will open a terminal at least once.

Korean-language usability

It runs Qwen3-ASR, whose official list of 30 supported languages includes Korean. Word-level timestamps (SRT) for Japanese and Korean use dedicated tokenizers, which the README calls out explicitly. Language can be auto-detected or set by hand. The model card publishes no separate benchmark for Korean.

Real constraints

  • Memory: the 1.7B model needs about 3.4GB of unified memory and the 0.6B about 1.2GB (quantized: 1.9GB and 0.5GB).
  • One file at a time, up to 4GB. Dictation runs 2 minutes by default and 10 at most.
  • Speaker labels are not in the docs. Intel Macs, Windows and Linux are out of scope.
  • Apache-2.0. No account, API key or cloud service.

Checked on 2026-09-14, Checked: From documentation. Written from the official docs and README. Not installed and tried by us.

Yap

Dictation on macOS that never leaves the device, built on Apple's Speech framework. 393From documentation

Pick this when. When you are on macOS 26 and all you need is menu bar dictation with no models to download and nothing to configure.

Installation

  1. brew install --cask frigadehq/tap/yap, or grab the .dmg from the releases page.
  2. Source build: git clone, then ./install.sh. Needs Xcode 26.
  3. No model to download, no API key, no account.

Korean-language usability

It uses Apple's SpeechAnalyzer and SpeechTranscriber directly and follows the system locale. There is no in-app language picker yet; the roadmap lists it as the likely next addition. Apple's docs do not publish a locale list and leave it to a runtime query, so whether Korean dictation works has to be checked on a Mac whose system language is Korean. That check is still outstanding.

Real constraints

  • macOS 26 (Tahoe) or newer, Apple Silicon only. No Intel Macs, no Windows.
  • Live microphone dictation only; file transcription is not in the docs.
  • There is no network code at all, so nothing leaves the device. History is stored locally in SwiftData. Free under MIT.

Checked on 2026-09-14, Checked: From documentation. Written from the official docs and README. Not installed and tried by us.

VoiceStudio

Local, open-source voice cloning, dubbing, and transcription with no account or API key 32,512From documentation

Pick this when. When you want transcription plus dubbing, voice cloning and audiobooks in one app, and can live with a heavy install.

Installation

  1. macOS (Apple Silicon): .dmg. The first launch needs a right-click, Open approval, and it then sets up a Python environment and downloads the default model.
  2. Windows 10/11: x64 MSI (pick the current-user build to install without admin rights). Linux: AppImage (glibc 2.39 or newer).
  3. Docker: the palashdeb/omnivoice-studio:stable image on port 3900. Images are linux/amd64 only.
  4. Source builds need Node 20 or newer (or Bun) and Python 3.11 or newer.

Korean-language usability

The default recognition engine is WhisperX, with 11 engines to choose from including Faster-Whisper, MLX Whisper, PyTorch Whisper and FunASR. Whisper-family engines support Korean, but the README only gives a language count (about 100) without naming Korean. Parakeet TDT and Moonshine are European-only and English-only respectively, so do not pick them for Korean files. We could not confirm a Korean UI.

Real constraints

  • Minimum spec is 8GB RAM and 10GB disk; recommended is 16GB or more and 20GB or more on SSD. With a GPU you need at least 4GB VRAM, and the larger engines need more.
  • Intel Macs cannot run the local Python backend (remote backend only).
  • Still in beta, AGPL-3.0. Downloaded models keep their own licenses; the default TTS model's weights are CC-BY-NC, which restricts commercial use.

Checked on 2026-09-14, Checked: From documentation. Written from the official docs and README. Not installed and tried by us.