Hold Right Option and speak. Release it, and the text appears at your cursor. The local model is 675 MB, and in local mode your audio never leaves your Mac.
Zero network requests in local mode · No account · No word limits
Real capture · a message dictated to an AI teammate · unedited output · Chinese and English in one sentence
macOS Dictation uses one recognition language at a time. Choose English and Chinese words are mangled. Choose Chinese and English terms are forced into Chinese. If you naturally mix both in one sentence, the built-in option breaks at the first step. Other apps introduce three more tradeoffs:
Many dictation apps send every clip to a remote server before returning text. That can include prompts, client names, and half-formed ideas.
Even apps with a local option often default to Whisper models with 1.6 GB or more in weights, competing with coding agents for the same memory.
Many speech models default to Simplified Chinese or mix character sets. That is how Simplified characters end up in a Traditional Chinese document.
Lamitype does one thing: turn your voice into text with minimal friction, privacy first.
The same sentence. Left: the Simplified Chinese most tools produce. Right: Lamitype's Taiwan Traditional Chinese, including the correct forms of 髮 and 乾.
Speech models can mix character sets. Lamitype applies deterministic OpenCC s2twp conversion, so the result uses Taiwan Traditional characters and vocabulary, including 軟體 instead of 軟件 and 影片 instead of 視頻.
The local model is 675 MB, less than half the size of Whisper large-v3-turbo. The MLX idle cache is capped at 1 GB, and one menu click unloads the model completely. Claude Code and Codex keep their headroom.
The model recognizes both languages inside the same sentence, with no engine or input-method switching. Say "這個 API 的 latency 有點高" and the English terms stay intact.
Local mode makes zero network requests, which you can verify with Little Snitch or a packet capture. For cloud recognition, your own API key connects directly to the provider, with no Lamitype server in the middle. The code is MIT licensed, so you can inspect the privacy claims yourself.
Free, no account. A short setup guide walks you through microphone and accessibility permissions.
Any text field, in any app.
Your text appears at the cursor.
Lamitype can also proofread without adding another bundled model. Select text in any app, double-tap Right Option, and Text Polish fixes typos, grammar, and punctuation in place. Every change appears in a diff, so you can see exactly what changed.
The honest limit: it is a proofreader, not a rewriter. It is designed to preserve meaning, tone, and mixed-language wording. Text Polish uses Apple Intelligence on device and requires macOS 26 or later. Its edits are conservative, and sometimes it decides nothing needs changing.
Real capture · a repeated phrase is tightened · every edit appears in the diff · the demo is in Chinese, but Text Polish works the same way in English
Bars compare model-weight sizes at each tool's default precision. Lamitype's local Qwen3-ASR 0.6B model is 675 MB after 4-bit MLX quantization. A 464 MB 4-bit Whisper Turbo model exists, but its Chinese output tends toward Simplified. The claim is therefore not smallest overall, but lightest among the models here that produce natural Taiwan Traditional Chinese. With a cloud engine and your own OpenAI or Gemini key, local model memory drops to zero.
Completely free under the MIT license. No subscription, no word limit, no account. Optional cloud dictation uses your own API key, and the model provider bills you directly. Nothing routes through Lamitype.
In local mode, turning speech into text makes zero network requests, and you can verify that with Little Snitch or a packet capture. Audio only leaves the machine if you enter your own API key and switch on cloud recognition, and then it goes straight to that provider.
The model download is 675 MB, and it occupies about 2.1 GB of RAM while loaded, with a 1 GB cap on the MLX idle cache, so an all-day session does not climb the way an unbounded transcription process does. When you are done, one menu click unloads the model and returns the memory.
Yes. It handles Chinese and English in a single sentence, such as "這個 API 的 latency 有點高," without switching engines or input methods.
Most models train on a mix of Simplified and Traditional text, so the character set you get is whatever the decoder found more probable that second. Lamitype converts the output deterministically with OpenCC s2twp, so the character set is not decided by the audio. It is always Taiwan Traditional, with Taiwan vocabulary.
All Apple Silicon Macs with M-series chips running macOS 15 or later are supported. Intel Macs are not supported because the local engine depends on Apple's MLX framework.
They are the same product. HushType was renamed Lamitype in 2026. The developer, codebase, and version history are unchanged and visible on GitHub.
Try it free. No account, no subscription, nothing to cancel.