Runs entirely on your Mac
Speak the way you actually talk. Get writing you’d send.
Hold a key, speak, release. Dictation that survives two languages in one sentence, on-device, for a one-time price.
Code-switching
Two languages in one sentence, and the text still reads.
Same utterance, both sides. Left is what was actually said, right is what landed at the cursor. Nothing here is retyped, and nothing is translated.
Spoken, disfluencies and all
uh, pwede ba natin i-move ‘yung meeting to Thursday? kasi, um, wala si Maria
With cleanup on — the language mix untouched
Pwede ba nating i-move ‘yung meeting to Thursday? Wala kasi si Maria.
By default you get exactly what you said, in the language you said it — no cleanup at all. Turning cleanup on removes the “uh” and fixes the punctuation. It does not translate, and it does not pick one of your languages over the other.
When you do want English, ask for it by voice Go
Select the Taglish you just dictated, hold the key and say “translate this to English”. It is rewritten in place. Translation is not a setting you leave on — the spoken instruction is what triggers it, so the same hotkey never silently changes your language. Advertised for Taglish into English only; the other direction is not good enough yet.
Measured, not claimed
We tested it against the dictation already on your Mac.
Thirty everyday Taglish sentences — work chat, email, notes — recorded once and run through both. Same audio, same day.
Apple’s dictation
Lost the Filipino in 28 of 30 sentences.
Bunso
Lost it in none.
Two of the thirty, copied exactly as each one produced them:
What was said
Okay lang ba kung sa Zoom na lang tayo? Ang layo kasi.
Apple heard
Okay.
Bunso heard
Okay lang ba kung sa Zoom na lang tayo? Ang layo kasi.
What was said
Bakit ang init ngayon? Parang wala yung aircon sa office.
Apple heard
Bakatam, Unique, and Paramvallayum, Aircon, office.
Bunso heard
Bakit ang init ngayon parang wala yung aircon sa office?
How this was tested
Thirty sentences recorded once, written down by hand first, then given to both without retries or edits. Nothing was cherry-picked to flatter us: every sentence counted, including the ones we got wrong. Bunso is not perfect on Taglish — it drops a hyphen or a question mark here and there, and you can see that above. It does not lose your sentence. Measured 14 August 2026 on an Apple Silicon Mac.
The same hotkey behaves differently per app.
Polished prose where prose belongs. Verbatim where a stray comma breaks the build.
Measured, on one machine, on one day.
| What was measured | Result | Note |
|---|---|---|
| Default speech model, end of speech to text | ~1.1 s median | English dictation, short utterances. This is the model you get in English-first locales. |
| Multilingual model, end of speech to text | ~2.4 s median | The one that survives a sentence in two languages — and the default in the Philippines and much of Southeast Asia. If code-switching is why you are here, this is your number. |
| Cleanup pass | 392 ms p50 | Added on top of transcription. Worst case in that run was 1.5 s. |
| Cleanup quality against a cloud service | 17/18 vs 18/18 | The cloud won, at 2.7× the time. Stated here rather than in a footnote. |
Speech timings measured on Apple Silicon, August 2026. The cleanup figures were measured 2026-08-09 on an idle machine, benches run back to back. Exact machine to be confirmed before launch. Your numbers will differ.
Your speech never leaves the machine.
There is no account and no telemetry, and transcription happens on your Mac. Your audio never leaves it, under any setting. The table below is the full list of things that can ever cross the network, and what governs each one.
| Data | Leaves your Mac? | Governed by |
|---|---|---|
| Audio you dictate | Never | No setting exists that would send it. |
| Transcribed text | Never, by default | Only if you turn on cloud cleanup with your own key. |
| Cloud cleanup (Go, optional) | Text only, if enabled | Off by default. Goes to the provider whose key you entered. Audio is never sent, under any setting. |
| Custom vocabulary | Never | Stored in your app support folder. |
| Tasks, if you mirror them to Apple Reminders (Go, optional) | To Apple, if enabled | Off by default. Bunso writes to the Reminders app on your Mac; if your Reminders sync to iCloud, your tasks sync with them, under Apple’s terms rather than ours. |
| Model files | Downloaded once | First launch, or when you add a language. |
| Update check | Version number only | Only when you ask it to check. |
| Buying a license | Your email address | Your payment provider tells our server you paid, and the server emails you a key. We keep your email, the key as a hash, and the payment reference. Two companies see your address: the payment provider, and the service that sends the email. Neither is used for anything else, and there is no mailing list. |
| License activation | Once per Mac | One request when you activate a Mac, carrying your key and a hashed identifier for that Mac. After that your license is verified on your Mac, offline, forever. |
This mirrors the full privacy policy row for row.
Pay once, or don’t pay.
No subscription. Updates included.
Free, unlimited, permanent
$0
Not a trial. No daily cap, no watermark, no nags. The same speech models and the same accuracy as Go — nothing about recognition quality is held back.
- Unlimited dictation, no cap
- Code-switching and Taglish, including the multilingual model
- The local cleanup pass
- Custom dictionary and tone presets
- Per-app behaviour: verbatim in a terminal, polished prose in Mail
- Learns from your corrections
- Foot pedal, latching and hands-free control
- Your last 10 dictations, with up to 5 saved keywords
Go
one-time$19
Everything in Free, plus six things. That is the whole list.
- Bring-your-own-key cloud cleanup — your provider, your bill, off by default. Audio is never sent, under any setting
- Voice editing — select text, speak an instruction, it is rewritten in place. Asking for a translation counts as an instruction
- Extra hotkeys that force a mode when the per-app setting is wrong
- Spoken notes and to-dos — a key that writes down what you said instead of typing it. Your first words decide: “remind me to pay the loan on Friday” becomes a task with the time set; anything else becomes a note
- Those tasks mirrored into Apple Reminders, in a list you already keep, so they reach your phone
- Unlimited history and unlimited saved keywords
Coming soon In active development
- Action keywords — say “send an email to Maria about Thursday” and a compose window opens, pre-filled. It prepares actions; it never sends, submits or pays for anything by voice.
What it doesn’t do.
Apple Silicon only
Intel Macs are not supported and will not be. The speed depends on the neural engine.
A few gigabytes on first run
The app ships small. The speech model and the cleanup model are fetched once, about 2.6 GB together — or roughly 0.5 GB if you decline the cleanup pass.
Three permission grants
Microphone, accessibility and input monitoring. Typing into other apps is not possible without them, and they are why Bunso cannot be sold through the Mac App Store.
A slow first launch
The speech model is compiled for the neural engine the first time you run it — a few minutes, and again after each update. After that, starting up is a short warm-up during which hotkey presses queue rather than fail. Not yet measured on the shipping app.
Questions people actually ask.
Does it work without an internet connection?
Yes. Speech recognition and the cleanup pass both run on your Mac, and out of the box nothing leaves it. There is one optional exception, off by default: you can supply your own API key and have the cleanup step run at a provider you choose.
Is my audio ever uploaded?
No. Audio is never sent anywhere under any setting. That claim is absolute, and the privacy page lists every other thing that can be transmitted and what governs it.
Which Macs does it run on?
Apple Silicon Macs, on macOS 15 or later. Intel Macs are not supported and will not be — the speed depends on the neural engine.
How big is the download?
The app ships small. The speech model and the cleanup model are fetched once, about 2.6 GB together — or roughly 0.5 GB if you decline the cleanup pass.
Which permissions does it need, and why?
Microphone, accessibility and input monitoring. Typing into other apps is not possible without them, and they are why Bunso cannot be sold through the Mac App Store.
Can it handle two languages in one sentence?
That is what it is built for. The multilingual model transcribes English and Filipino mixed mid-sentence without translating either one. Real transcripts from a thirty-clip test, and the same clips through the dictation already on your Mac.
Is the free version limited?
Not on dictation. It is unlimited, with no cap, no trial and no nags, and it uses the same speech models at the same accuracy as the paid version. Nothing about recognition quality is held back.
What happens if my license lapses?
Cleanup moves back onto your Mac; it never switches off. A lapsed license also never deletes anything you wrote down — not a saved keyword, not a note, not a to-do, and not a reminder already copied into Apple Reminders. It only stops you adding new ones.
Try it on your own sentences.
The free tier is unlimited. Buy Go when the extras earn it.