mybunso.app

Taglish dictation on a Mac

Examples measured 10 August 2026.

Most dictation tools do not handle two languages in one sentence badly. They do not handle it at all.

The problem is the middle of the sentence, not the language

Dictation software asks you to pick a language. That works if you speak one. If you are Filipino, or Singaporean, or Indian, or speak Spanish at home and English at work, the unit that carries your meaning is not a language — it is a sentence with two of them inside it:

Na-deploy ko na yung latest build sa staging kanina.

Pick English and the recogniser tries to fit Filipino sounds to English words. Pick Filipino and it does the reverse to build and staging. There is no third setting, because the mixing happens mid-sentence and a language setting is chosen before you speak.

The hardest part is a word that is both

Words like na-deploy, i-review, pakicheck and mag-Friday are an English root under a Filipino affix. They are not slang and they are not a mistake — that construction is how the grammar borrows a verb, and it is completely ordinary in professional Filipino writing.

They are also the exact point where a recogniser with no Filipino gives itself away. It either drops the affix, or it invents a word that sounds close and means nothing. A tool can look fluent on your English and still be unusable, because everything holding your sentence together is gone.

What that looks like in practice

Six of the thirty recordings, chosen to cover the different ways people write at work. This is what the multilingual speech model produced, before the cleanup pass runs.

Six of the thirty test recordings, verbatim.
What was said What Bunso heard
Na-deploy ko na yung latest build sa staging kanina. Na-deploy ko na yung latest build sa staging kanina.
Kailangan natin i-refactor yung authentication middleware bago mag-Friday. Kailangan natin i-refactor yung authentication ng middleware bago mag-Friday.
Send mo na yung invoice kay Ma'am Divina, tapos i-cc mo si accounting. Send mo na yung invoice kay Ma'am Divina tapos isi-same mo si accounting.
Set natin yung call sa Monday para ma-discuss yung budget. Set natin yung call sa Monday para ma-discuss yung budget.
Bakit ang init ngayon? Parang wala yung aircon sa office. Bakit ang init ngayon parang wala yung aircon sa office?
Nag-crash yung app pag nag-open ka ng settings tapos nag-rotate ka ng screen. Nag-crash yung app pag nag-open ka ng settings tapos nag-rotate ka ng screen.

Read the third row closely, because it is a failure and it is ours: i-cc came back as isi-same. Two letters spoken as letters, inside a Filipino affix, is genuinely hard and we did not get it. The full thirty, with the built-in macOS dictation alongside for comparison, are on the head-to-head page — every clip, including the ones we lost.

Hyphens are the residual problem, and they are fixable

Once the speech model actually knows the language, most of what is left is not recognition at all — it is spelling convention. The model hears ma-save correctly and writes masave; it hears i-share and writes ishare. The word is right and the hyphen is missing.

That is a rule, not a model problem, and Bunso applies one: it splits known Filipino affixes off English roots. The words in your custom dictionary are the list it splits against, which is why adding your own vocabulary helps here even though the dictionary's more obvious job is to prime the recogniser.

Being honest about the limits

What the app does with the text once it has heard it, or the full thirty against macOS dictation.