I'm Sona Orbit, the solo developer behind Parroting. Parroting is a free Mac dictation app built on the open-source Handy (MIT license). Features and menu names here refer to Parroting 0.4.19. Prices for other tools come from my Typeless alternatives comparison, checked on October 31. Prices change, so check each tool's website.
1. First, choose a voice typing setup for coding
Talking to an AI coding tool is different from replying to a message. You mix in English terms, project names, and commands. You also think out loud and change your mind mid-sentence. Start by comparing these four options:
| Approach | Examples | Where speech is recognized | Filler and false-start cleanup | Cost |
|---|---|---|---|---|
| Local recognition + your choice of AI | Parroting, Handy, VoiceInk | On your Mac | Connect an AI service you choose (Parroting also supports Apple Intelligence) | Parroting and Handy are free; VoiceInk starts at $25 as a one-time purchase |
| Cloud subscription | Wispr Flow, Typeless | In the cloud | Built in | Free tiers allow about 2,000 words per week; Pro costs about $15–$30 per month |
| Local models + paid features | Superwhisper | Locally or in the cloud | Custom modes | From $8.49 per month, or $249.99 for lifetime access |
| Built into the system | macOS Dictation | Some languages run locally; Chinese (Taiwan) is not on Apple's on-device list | None; verbatim output | Free |
If you dictate thousands of words to AI every day, a free tier of about 2,000 words per week runs out quickly. Parroting recognizes speech locally with no word limit. You only need to connect an AI service if you want speech cleanup. You can use Apple Intelligence or a free Gemini key. For a fuller comparison, see Typeless alternatives.
2. Get technical terms right when you mix Chinese and English
When coding in Chinese, you might say: 「幫我看一下 webhook 那支 handler,為什麼收到 payload 之後沒有寫進 Redis?」 (Can you check that webhook handler? Why doesn't it write to Redis after receiving the payload?) Getting a mixed-language sentence like that right takes two steps.
Step 1: Choose a speech model
Choose a model in “Settings → Talk → Speech model & language.” Select “Manage models” to download others. Here are three starting points:
- SenseVoice Small (about 250 MB): The fastest option. Parroting marks it “Recommended,” and it runs smoothly on older Macs.
- Qwen3-ASR: The fewest Chinese character errors of these models. The 1.7B model is about 1.5 GB and slower; 0.6B is a middle ground.
- Breeze-ASR-25: Tuned by MediaTek Research for Taiwanese Mandarin with English mixed in. It is the slowest, but handles Taiwanese phrasing better.
Parroting's Chinese formatting adds full-width punctuation and spaces between Chinese and English. Traditional or Simplified Chinese output follows your setting. You do not need to add spaces around English terms yourself. For model sizes, speed, and Chinese character error rates, see the Mac local speech recognition model comparison. To find a model that suits your accent, use “Read aloud” under “Personalization.” Read 10 sentences in about 2 minutes, then check the character error rate. Switch models and read another round to compare.
Step 2: Use the dictionary for project names and abbreviations
A general-purpose model will not know your project's code names or a new tool's name. Parroting applies dictionary corrections after recognition, so they still work when you switch models.
- Correct it once, then use it in the next sentence: If you fix a misheard word after it is pasted, such as changing “Jmini” to “Gemini,” Parroting records the original and corrected text under “Personalization” for your review. Select “Add to dictionary,” and the correction works from the next sentence.
- Rewriting a whole sentence will not pollute the dictionary: Parroting uses pronunciation to tell a misheard word from a change in wording. Reworded text does not go into the dictionary or training data. You can correct its decision if it gets it wrong.
- Import a batch of terms: In “Settings → Words & history → Dictionary,” select “Add” or import a CSV. Put the correct spelling in the first column and the misheard version in the second. You can add table names and service names together. When AI cleanup is on, Parroting also sends the dictionary to the AI so it does not change those terms by mistake.

3. Clean up false starts for free
When explaining a complex change to AI, it is hard to say everything cleanly in one pass. It often sounds like this:
「呃,幫我把那個 login 的 API 改成……不對,先不要改 login,先去檢查 middleware 裡面的 token 驗證,過期的話回傳 401,然後順便記一筆 log。」 (Uh, change the login API to... no, don't change login yet. First check token validation in the middleware. Return 401 if it has expired, and log it too.)
If you paste that verbatim, you still have to delete “uh,” “no,” and the abandoned request to change login. Parroting's AI speech cleanup removes filler and false starts before pasting. It only cleans up what you said. It does not add new instructions. Without an AI service configured, Parroting uses local rules for basic cleanup.
There are two ways to do this without a monthly subscription:
- Apple Intelligence: Select Apple Intelligence under “AI Services.” You do not need a key, and the text stays on your Mac. This requires an Apple silicon Mac running macOS Tahoe (26) or later, with Apple Intelligence enabled in System Settings.
- A free Gemini API key: Follow the three-minute setup guide to get one from Google AI Studio. You do not need to add a credit card. The free tier has per-minute and daily limits, but they are usually enough for ordinary dictation and rewrites. If you hit a limit, Parroting pastes the locally cleaned version so you can keep working. Google says content sent through the free tier may be used to improve its products. If you often dictate confidential work, consider a paid plan or Apple Intelligence instead.
To turn every dictated passage into bullet points, select “Organize” in the rewrite panel and turn on “Auto replace.” Parroting will then paste the organized version directly, even when the panel is closed. For costs, see what it costs to use your own API key.

4. Two ways to dictate: hold to talk or tap to keep listening
The default shortcut is right ⌥ + right ⌘, within reach of your right thumb. You can use it in two ways:
- Short commands, hold to talk: Put the cursor in Cursor, a Claude Code terminal, or a commit message field. Hold the shortcut while you speak, then release it to paste at the cursor. Press Esc to cancel before you finish.
- Long passages, tap to keep listening: To dictate a refactoring plan while looking at code, tap the shortcut and release it within 0.3 seconds. The floating window stays open and keeps listening. At each pause, Parroting cleans up and pastes a segment. Tap again when you are done. You can switch apps while speaking. Text goes to the field you last typed in.
5. Troubleshoot pasting, shortcuts, and Bluetooth audio
A developer's desktop can have external displays, several terminals, an IDE, Raycast, and Karabiner-Elements. If something goes wrong, check these first:
Text is recognized but does not paste into Cursor or the terminal
- Accessibility permission: Parroting needs Accessibility permission to paste into other apps. Check that Parroting is enabled in “System Settings → Privacy & Security → Accessibility.” If pasting stops right after an update, turn the permission off and back on.
- The app does not report the cursor position: Codex and other Chromium-based apps sometimes do not report where the cursor is. Since 0.4.14, Parroting asks again. If there is still no answer, it pastes anyway and leaves the text on the clipboard so you can paste it yourself. Update to the latest version first.
- The cursor is not in a text field: Pasting will not work if focus is on a button or list. The text is not lost. Open “History” to copy what you just said.
The shortcut does nothing or triggers another app
Raycast, Alfred, Karabiner-Elements, or a Hyper Key setup may be using the right-side modifier keys. Choose an unused combination in “Settings → Talk → Shortcuts.” If the menu bar shows “Shortcuts blocked by Secure Input,” another app has temporarily locked keyboard access. Password fields and terminal secure input are common causes. Leave that field to restore the shortcut.
AirPods miss the first two words, or music sounds muffled
This is a common macOS issue when using a Bluetooth headset microphone. Starting a recording requires a Bluetooth channel switch, which can cut off the beginning. Music in the headset also temporarily drops to lower quality. At your desk, choose your Mac's built-in microphone or a desktop USB microphone under “Settings → Talk.” Leave the AirPods for listening. This makes it less likely that the start of a sentence will be missed.

6. Frequently asked questions
When I dictate to Claude Code or Cursor, does my code or audio get sent out?
Recording and speech recognition happen on your Mac. Audio is not uploaded. Only when you enable AI cleanup or rewriting is the recognized text sent directly to the AI service you choose. It does not pass through me. With Apple Intelligence, the text stays on your Mac too.
Can I use my OpenAI, Anthropic, or OpenRouter key, or a model I run locally?
Yes. Alongside Gemini and Apple Intelligence, AI services include OpenAI, Anthropic, Groq, OpenRouter, and custom endpoints. If your local model provides a compatible endpoint, enter its URL and model ID.
Should I turn on tone hints while coding?
You usually do not need them for instructions to AI. When a hint appears, select “Mute hints in this app” to stop hints in your terminal or IDE. The local tone model can still show hints when you switch to Slack, LINE, or Mail to reply to coworkers or clients.
What if my external keyboard has no right ⌥ or ⌘ key?
Choose a shortcut you can reach in “Settings → Talk → Shortcuts.”
Further reading: Choosing a local speech recognition model for Mac, Fixing voice typing errors, and Comparing Typeless alternatives.
