Sageio Type
On voice, typing, and getting words right
Notes on dictation, Asian-language voice input, and turning what you say into text you'd actually send.
部落格文章目前僅提供英文版。
Cantonese voice input: the gap between spoken and written
Cantonese is spoken one way and conventionally written another. A speaker dictating 我唔係好清楚 faces a real choice most tools never acknowledge: keep the colloquial form for chat, or render Standard Written Chinese for formal work. Here's why the diglossia, traditional characters, and Hong Kong vocabulary make Cantonese voice input its own problem.
中英混講的語音輸入：亞洲職場最真實的難題
「這個 sprint 我們先把 deck 做完」——英文詞嵌在中文句子裡，是華語科技與商務圈的工作語言，不是講話隨便。可是鎖定單一語言的語音輸入，總會把其中一種絞爛。這篇談中英混講到底難在哪，以及好的處理該長什麼樣。
Dictating email that sounds like you wrote it
Speaking an email and writing one are not the same act. This is why most voice typing produces a transcript instead of a message, and what a dictation tool has to add on top of your words — greetings, paragraphs, and tone that matches where the text is going.
Dictation as an accessibility tool: RSI, fatigue, and voice-first workflows
For people who can't type comfortably all day — repetitive strain, chronic pain, fatigue — dictation isn't a productivity toy, it's the primary way into the computer. What a voice-first Mac workflow really looks like, and why accuracy and a clean rewrite pass matter more here than for a casual user.
Drafting long-form by voice: from ramble to readable
Talking a document into existence produces structure-free text, and that's fine — the rewrite pass is what gives it paragraphs and shape. Here's why dictating a long draft, then editing what you already said, is a lighter task than typing into a blank page, and how to work with the raw material voice gives you.
Voice-typing Slack messages: casual tone is harder than it looks
Dictating into a chat channel isn't the same as dictating an email. The output that reads clean in a document reads stiff and wrong in Slack. Here's why casual register is its own hard problem, and why the rewrite that polishes an email has to dial itself down to fit a chat window.
Why your dictation tool needs to learn your vocabulary
A generic speech model has never heard your colleague's name, your product's codename, or the acronym your team says ten times a day — so it guesses a common word that sounds close, every time. A personal dictionary turns a repeated correction into a one-time lesson: teach the tool your terms once, and it stops mishearing them.
Who puts the commas in? Punctuation and formatting in modern dictation
Old dictation made you say the punctuation out loud — 'comma', 'new paragraph', 'period' — and the words appeared literally when it guessed wrong, breaking your train of thought. Modern AI-pass dictation infers punctuation, paragraph breaks, and lists from rhythm and meaning, so you just talk. Here's how that works, and why East Asian punctuation is its own harder problem.
From sound to sentence: how AI dictation actually works
Modern dictation isn't one step, it's two: a speech model turns sound into a raw transcript, then a language model rewrites that transcript into finished text — punctuation, removed filler words, and a tone that fits where you're writing. Understanding the two stages explains why some voice typing reads like a court transcript and some reads like you wrote it.
How to dictate on a Mac in 2026: options, trade-offs, and what changed
There are three broad ways to dictate on a Mac — the option built into the operating system, browser-based tools, and dedicated apps — and they no longer differ only in accuracy. This is a neutral survey of what each category does well, where each one costs you, and how an AI rewrite pass quietly changed the whole workflow from raw transcription to finished text.
在 Mac 上用中文語音輸入，為什麼常常不好用？
中文是全世界母語人口最多的語言，可是在 Mac 上用語音輸入中文，失敗率往往比英文還高。問題不在麥克風，而在於中文的同音字太多、字與字之間沒有空格——把說出來的話變成你真正會送出去的文字，需要先聽懂整句話，這正是多數工具跳過的一步。
Thai voice typing: a script with no spaces is a segmentation test
Thai is written with no spaces between words — spaces mark only phrase and sentence boundaries. So turning Thai speech into correct text is inseparable from word segmentation: the model has to decide where each word begins and ends, and Thai also carries tones and gendered polite particles (ครับ / ค่ะ) that a register-aware rewrite has to get right.
繁體中文語音打字：沒人談的「簡體輸出」問題
多數語音模型吐出來的是簡體中文，台灣和香港的使用者拿到的，是自己根本不會這樣寫的文字。真正的繁體支援是詞彙層級的（軟體 vs 軟件），不是把字換一換就好。這是一個建模選擇，不是字型設定。
Vietnamese voice typing: six tones and a diacritic minefield
Vietnamese is written in a Latin script loaded with tone marks, and one mark changes a word entirely — ma, má, mà, mả, mã, mạ are six different words. Typing those diacritics is slow and fiddly, which is exactly why voice input matters so much here, and why regional accents make it hard. A look at dictating tiếng Việt on a Mac.
Voice input for multinational teams: write in the language the reader needs
The teams that run cross-language meetings also spend their day writing cross-language email and chat. Voice input lets a non-native writer speak in whichever language comes fastest and get back polished text in the register — and language — the reader actually needs.
語音打字完整指南：在 Mac 上用說的寫中文，2026 年有什麼不一樣
在 Mac 上用語音打字寫中文，這幾年真正變了的地方，不是辨識更準，而是多了一道「潤飾改寫」的步驟——把你說出來的口語，整理成你會直接送出去的文字。這篇用中立的角度，說明現有的幾種做法各自適合誰，以及怎麼判斷你需不需要更進一步的工具。
Turning spoken thoughts into meeting notes, right after the call
The minutes right after a call are when decisions and action items are sharpest in your head. Instead of typing them up, debrief by voice and let dictation shape the recap into notes: decisions, owners, follow-ups, a draft email. Here is why speaking your takeaways beats typing them, and how to make the habit stick.
Where does your voice go? A privacy checklist for dictation tools
A voice-typing tool hears everything you say — meetings, passwords read aloud, half-formed thoughts. This is a practical checklist for evaluating any dictation tool: what happens to the audio, whether your speech trains models, where history lives, what actually leaves your Mac, and which controls you get.
Chinese dictation on a Mac: why voice typing breaks on the world's biggest language
Chinese has more first-language speakers than any other language, yet voice typing fails on it more often than on English. Tones and a huge homophone space mean Chinese dictation needs sentence-level context, not word-by-word transcription — plus the IME tax, traditional-vs-simplified output, and Chinese-English code-switching. Here's what actually makes Chinese dictation usable.
Dictating English as a non-native speaker: accuracy beats speed
Most voice typing is tuned to a narrow set of native accents and stumbles on the English the rest of the world actually speaks. Here's why accent robustness and a quiet grammar-smoothing rewrite matter more than raw speed for second-language dictation.
How accurate is AI dictation? A framework for judging it yourself
Accuracy in dictation isn't one number. Raw word accuracy, formatted-output accuracy, and whether the tool kept your meaning are three different things — and a language count tells you almost nothing about per-language quality. Here's how to test any dictation tool on your own speech in a few minutes and read the result honestly.
Japanese dictation on a Mac: kanji conversion is the real test
Turning Japanese speech into kana is the easy half. Choosing the right homophone kanji — 公開, 後悔, 航海 all sound like koukai — needs the whole sentence, and matching です/ます against plain form matters the moment you dictate an email instead of a chat. Here's what separates Japanese dictation you can send from Japanese dictation you have to rewrite.
Korean dictation: spacing and honorifics make or break it
Korean word spacing (띄어쓰기) is genuinely hard and instantly visible when it's wrong, and the same sentence changes shape depending on whether you're writing a Slack DM or an email to a director. That's why Korean dictation needs a rewrite pass over the raw transcript, not just transcription — plus the mixed English loanwords that run through every business sentence.
Dictating in two languages at once: the code-switching problem
Real Asian-workplace speech mixes languages mid-sentence — English product terms inside Chinese clauses, katakana loanwords in Japanese. Single-language dictation forces a pick and mangles the other. Here's what good mixed-language handling looks like: one clean sentence, each language written as itself.
Traditional Chinese voice typing: the simplified-output problem nobody talks about
Most speech models emit simplified Chinese, so Taiwan and Hong Kong users get text they'd never write. Real traditional support is vocabulary-level — 軟體 vs 软件, 資訊 vs 信息 — not a font setting or an afterthought conversion. Here's why it's a modeling choice made per user.