Voice dictation on Mac for Uzbek and Kazakh
Voice dictation on Mac works in any app where your cursor sits: press a key, speak, release, and the text appears. It's faster than typing for long messages, drafts, and AI prompts. The catch is almost always the same: everything depends on which language you're speaking.
How voice typing works on a computer
When you hold a hotkey and speak into the microphone, three things happen in sequence. The app records a short audio clip, passes it to a speech recognition model, and inserts the resulting text wherever your cursor is. Release the key and recording stops.
Here's why this differs from dictating on a phone. A mobile keyboard only recognizes speech inside its own input field. Computer dictation works across every application: browser, messenger, document, terminal. If there's a cursor, text lands there.
macOS system dictation and Windows voice typing do the same thing, with one hard limit: the language list is fixed. Central Asian Turkic languages don't appear on those lists.
Why does system dictation fail on Uzbek and Kazakh?
Global engines trained primarily on English. Uzbek and Kazakh are agglutinative: a single word carries the meaning of a full sentence in other languages, and each word takes dozens of forms. Publicly available training data for these languages is scarce, and pronunciation varies significantly across regions.
The result: global engines exceed 30% word error rate on Uzbek. Every third word comes out wrong. Dictating at that quality produces noise, not text.
A model trained specifically for Central Asian languages errors on roughly 9% of words on Uzbek, one word in eleven. That's more than a three-times improvement, and you feel it immediately: instead of garbled output you get readable text.
There's a nuance for Kazakh. The model is also specialized for the region on Kazakh and Kyrgyz, but no separate measured WER figure for those languages has been published. If Kazakh accuracy is your priority, the Kazakh speech to text article goes deeper.
What works well for dictation, and what's better typed
Voice input wins where your thought outruns your fingers. It loses where symbol-level precision matters.
Good to dictate:
- Messages in chats and work emails
- Text drafts and notes
- Prompts for Claude Code, Codex, and other AI tools
- Task comments in project trackers
- Terminal commands you know well
Better typed out:
- Code: one wrong letter changes meaning, and a 9% word error rate turns a function into garbage
- Formulas, addresses, URLs
- Tables and structured data
The AI-prompt scenario deserves a specific note. When you're describing a task to Claude Code or dictating an instruction into Codex, natural-speech accuracy is enough. A prompt, like an email, tolerates one extra word in eleven. The code the model writes back you'll review by hand anyway.
| Scenario | Good fit for dictation? |
|---|---|
| Message to a colleague in chat | Yes |
| Client email, 3-5 paragraphs | Yes |
| Prompt for Claude Code or Codex | Yes |
| Comment on a task in Jira or Notion | Yes |
| Terminal command (familiar one) | Yes |
| Python or TypeScript function | No |
| SQL query with table names | No |
| Spreadsheet cells with numbers | No |
If you need to transcribe someone else's audio rather than dictate your own text, that's a different task: see the Uzbek speech to text article for that workflow.
How to set up dictation so it stays out of your way
One practice covers most of it: assign a single key under your thumb and stop thinking about it. Hold it down, speak the whole phrase, release. Speak in connected phrases, not individual words: the model produces cleaner output when it has sentence context to work with.
Before speaking, check where your cursor is. That's the only way to misfire: text drops into whatever field was last active. Click the right place, then hold the key.
The dictated result is worth a quick scan, but for a first draft it holds up fine.
Dictation works the same way in a chat window, a document, and a terminal. That's the point of system-wide voice input: one gesture, any window.
Does offline dictation actually change anything?
It does, for two reasons. One is connectivity. The Sezish 225 MB model runs entirely on your Mac: internet is only needed to download the app. After that you can dictate on a plane, at a cabin, or on a corporate network that throttles external traffic.
The second reason is what happens to your words. Prompts to a terminal, task descriptions for colleagues, letters to clients: all of that carries context you may not want leaving your machine. With local processing, no byte of audio exits the laptop. The details are in the privacy policy.
Honest limitation: fully offline dictation works on Mac. The Windows version is an early beta with dictation only, without the full feature set of the Mac version. On mobile, processing goes through a server over TLS, and audio is not stored after the response.
Download for Mac is available directly on the site. The app lives in the menu bar and also handles meeting recording with transcripts in the same five languages.
FAQ
Does dictation work without an internet connection?
On Mac, yes. The 225 MB model is installed locally and runs without a network connection. No internet is required during recording or processing. Mobile apps process audio through a server and require a connection.
Does what I dictate get sent to an external server?
On Mac, nothing leaves your computer. All processing happens locally. When using mobile apps, audio is sent over TLS, processed in memory on the server, and not stored after the response. Recordings are not used to train models.
Can I dictate into a terminal or into Claude Code?
Yes. Dictation works in any active field: terminal, a browser tab with Claude Code, a Codex window, it doesn't matter. Prompts and task comments are a natural fit for voice. Writing the code itself by voice isn't: one swapped letter in a variable name breaks logic.
What happens if I mix two languages in one sentence?
Sezish recognizes five languages: Uzbek, Kazakh, Kyrgyz, Russian, and English. Both halves of a mixed phrase, say Russian and Uzbek in one sentence, fall within that supported set. No separate accuracy figure for mixed speech has been published, so reread such a phrase before sending it.
How accurate is dictation for Uzbek and Kazakh?
On Uzbek, the model trained specifically for Central Asian languages errors on 9% of words, one in eleven. Global engines on Uzbek exceed 30%. For Kazakh and Kyrgyz the model is specialized for the region; separate measured WER figures for those languages have not been published.
Is there a Windows version with the same dictation?
A Windows version exists, but it's an early beta. It supports dictation only: the Mac features, including meeting recording, are not available yet. For full offline dictation, you currently need a Mac.