使用 Reki 裝置端語音辨識

Reki note can finish transcription on your own device without the cloud. Windows offers “Reki on-device transcription,” where the app prepares everything, and Mac offers “Mac built-in transcription” using macOS speech recognition. This page covers both.

Last updated: September 30, 2026

What is on-device transcription?

On-device transcription finishes processing on your own PC or Mac instead of sending audio to a cloud AI service. This suits sensitive meetings or cases where you want to avoid sending data externally.

The concrete method differs by OS: Windows uses “Reki on-device transcription” and Mac uses “Mac built-in transcription.” Both are available under Settings > Transcription.

Windows: Reki on-device transcription

On Windows you can use “Reki on-device transcription,” a CPU-only multilingual local transcription method where Reki note prepares everything. There is no extra app to install and no server to start.

Everything that gets downloaded is managed by the app, so you can pick it and start using it right away. Audio and transcripts never leave your device.

There are two model packs. You can also upgrade to the multilingual pack later with an additional download.

Japanese & English (~1.1GB): base model only. Speaker separation and additional languages are not available. This is enough for Japanese and English.
Multilingual & speaker separation (~3.1GB): includes per-language routing and speaker labels. Recommended for meetings with multiple languages or when you want speakers separated.

Open Settings > Transcription in Reki note.
Under transcription method, select “On-device transcription.”
Choose the “Japanese & English” or “Multilingual & speaker separation” pack and start the download.
Components are prepared in order: Python runtime, speech engine, dependencies, then speech models. Depending on your connection, this takes a few to a dozen or so minutes.
When it shows “Ready,” setup is complete and it is used for subsequent meetings.

Mac: Mac built-in transcription

On Mac you can use “Mac built-in transcription” (called “lightweight transcription” in settings), which uses the speech recognition built into macOS. There is no model to download in the app, and processing stays on your Mac.

You can choose from three speech engines. If unsure, leave it on “Auto (recommended).”

Auto (recommended): uses the high-accuracy latest model when available, and falls back to the classic model otherwise.
High-accuracy latest model: uses the high-accuracy model (macOS 26 or later; requires a model download).
Classic model: uses the speech recognition engine built into macOS (macOS 13 or later).

Open Settings > Transcription in Reki note.
Turn on “Use lightweight transcription.”
Choose a speech engine if needed (Auto (recommended) is fine for most people).
If you chose “High-accuracy latest model,” press “Prepare model” to download it. Until it is ready, transcription uses the classic model.

How it differs from local Whisper

On Windows you can choose between on-device transcription and local Whisper; on Mac, between Mac built-in transcription and local Whisper. All of them keep audio on your machine, but they play different roles.

On-device transcription (Windows): Reki note prepares everything, so you just pick it. It runs on CPU only, with no GPU required, and includes automatic multilingual detection and speaker separation.
Mac built-in transcription (Mac): uses macOS speech recognition, so there is nothing extra to install. Lightweight and easy to use.
Local Whisper: a whisper.cpp-based method available on both Mac and Windows. You choose the model yourself, so you can fine-tune accuracy, size, and speed, and use a GPU for faster processing.

Start with the OS-native method (on-device transcription or Mac built-in transcription) and switch to local Whisper if you want to choose models in detail.

Troubleshooting

Install fails (Windows): check your network connection and retry. If it fails partway, the next attempt resumes the download.
Speaker separation is slow (Windows): it re-analyzes the entire audio and can take time. If you are in a hurry, transcribe without speaker separation.
Everything feels slow: turn off real-time transcription to reduce load during meetings and app startup.
Reclaim disk space: removing it from settings also deletes the downloaded models.

Recommended PC specs

On-device transcription works without a GPU, but comfort depends on your hardware. For CPU and memory guidelines and how it compares with local Whisper, see the “Machine specs for local Whisper” article.

If you are unsure whether your PC is fast enough, tell an AI chat about your PC specs for a quick check.