How to Dictate into Any Mac App with a Fully Local Speech Model

Tap one key, speak, and put the transcript at the cursor in any Mac app. The speech model runs locally, and you choose which model hears you.

Yes: you can use a speech-to-text model as a fully local, system-wide dictation tool on an Apple Silicon Mac. Rapid-MLX Desktop listens only during a dictation session, transcribes through a model running on your Mac, and pastes the result wherever the cursor is. It works outside the Rapid window in editors, browsers, chat apps, terminals, and most other places that accept paste.

The default trigger is a tap of Right Command. You can switch it to Right Option. Tap once to start recording, speak, then tap the same key again to stop. A small heads-up display shows the recording state while the local model works.

What Rapid-MLX adds to macOS

There are two separate jobs in voice input:

1. A speech model turns audio into text.
2. A system-wide utility captures the microphone, runs the model, and inserts the text into the active app.

A speech model alone only solves the first job. Rapid-MLX Desktop supplies the second: a global hotkey, microphone capture, model lifecycle, a recording HUD, local transcription, clipboard-safe insertion, recent history, and a vocabulary list for names or technical terms.

Set up local dictation

1. Download Rapid-MLX Desktop and open it on an Apple Silicon Mac running macOS 14 or later.
2. Open Audio, then Speech to Text.
3. Choose a transcription model and download it. The model must be on disk before dictation can be enabled.
4. Allow Microphone access. Rapid records only while a dictation session is open.
5. Grant Accessibility access so Rapid can detect the global hotkey and paste into another app. Quit and reopen Rapid after granting it; macOS applies this permission at launch.
6. Turn dictation on. Keep Right Command, or change the trigger to Right Option.
7. Put the cursor in any app, tap the trigger once, speak, and tap it again.

The transcript is placed on the clipboard and pasted in one operation. That avoids character-by-character injection, which is slow and can be mangled by CJK input methods. If a secure field blocks synthetic paste, the text remains on the clipboard so you can paste it manually.

Which speech model should you choose?

Rapid is not tied to one recognizer. Its audio model catalog exposes several local STT families through the same interface.

For Mandarin mixed with English product names or code, start with qwen3-asr. For fast Chinese or Cantonese, try sensevoice. For a broad multilingual default, use whisper-large-v3-turbo. For English only, parakeet-v3 is the specialized option.

There is no universal winner. Accent, microphone, room noise, vocabulary, and how often you switch languages all matter. Try the same one-minute sample with two models before deciding.

What stays local?

The important boundary is straightforward:

• The model weights are downloaded once and stored on your Mac.
• Microphone capture runs only between the two hotkey taps.
• Audio is transcribed through the Rapid server on your Mac.
• Raw recordings are not saved by default. Audio archiving is a separate opt-in setting.
• Transcript history and vocabulary are local application data.
• Finished text is copied to the macOS clipboard, then pasted into the frontmost app.

The app still needs network access when it downloads a model and checks for software updates. "Local dictation" describes the speech-processing path; it does not mean the entire application is permanently disconnected from the network.

Why model choice matters for Chinese dictation

Whisper is useful because it covers many languages with one model, but it is no longer the only practical local option. Chinese dictation often includes English names, acronyms, code identifiers, and product vocabulary. That makes code-switching and vocabulary hints more important than a single headline accuracy score.

qwen3-asr is designed for Chinese/English code-switching and accepts vocabulary context. Rapid's vocabulary panel turns names you care about into those hints. sensevoice is much smaller and is purpose-built for Chinese, Cantonese, Japanese, Korean, and English. The right choice can therefore be a specialized Asian-language model rather than a larger Whisper checkpoint.

Honest limitations

System-wide dictation needs both Microphone and Accessibility permissions. Some password fields and hardened applications reject synthetic paste. The first run after loading a model may be slower than later warm runs, and larger models consume more unified memory.

Local also does not automatically mean accurate. Whisper, Qwen3-ASR, SenseVoice, and Parakeet make different errors. Proper nouns still benefit from vocabulary hints, and noisy rooms still benefit from a decent microphone.

Frequently asked questions

Does Rapid-MLX dictation work in every Mac app?

It works in most apps that accept normal clipboard paste, including editors, browsers, chat apps, and terminals. Secure fields or hardened apps may block synthetic paste; in that case Rapid leaves the transcript on the clipboard for manual paste.

Is my dictation audio uploaded?

No. After the initial model download, dictation audio is sent to the Rapid transcription server running on your Mac. Raw audio is not archived unless you explicitly enable audio archiving.

What is the best local speech-to-text model for Chinese on a Mac?

Start with qwen3-asr for Mandarin with English code-switching and vocabulary hints, or sensevoice for fast Chinese and Cantonese recognition. whisper-large-v3-turbo is the broad multilingual alternative. Test a representative recording because accent and vocabulary change the result.

Can Rapid-MLX use Whisper for dictation?

Yes. Rapid supports Whisper Large v3, Large v3 Turbo, Medium, Small, and Base aliases. It also supports non-Whisper models including Qwen3-ASR, SenseVoice, and Parakeet.