Bunso vs the dictation already on your Mac

Measured 10 August 2026. Thirty sentences, one speaker, one take each.

The same thirty recordings through both. Every output published, including the one where the built-in dictation heard a word we missed.

How this was run

Thirty everyday code-switched sentences — English and Filipino mixed inside one sentence, the way people actually write at work — were written down first, then recorded once each by a single speaker in a normal room on a normal microphone. No retries, no re-reads, no edits. The same audio files were then given to macOS dictation and to Bunso.

macOS dictation was driven by hand, one clip at a time, in its en-US locale. That is not a handicap we chose: there is no Filipino locale to select. The current dictation engine ships thirty locales and none of them is Filipino or Tagalog, and the older speech API offers Philippine-accented English but still no Tagalog. The best case available to a Filipino speaker on a stock Mac is an English recogniser being handed speech it has no model for.

The Bunso column is what its multilingual model produced, before the cleanup pass runs. That is the honest comparison — it is speech recognition against speech recognition. The cleanup pass, which is what fixes punctuation and capitalisation, would flatter this column if it were included.

Where the built-in dictation wins

Stated first, because it is the part you cannot get anywhere else.

• It heard a word we missed. In one clip of the thirty, the reference is "Nag-fail yung CI kasi may lint error sa bagong component." macOS dictation got lint. Bunso heard length. On a corpus this small that is one clip, but it is a real recognition loss and no cleanup pass recovers it — a plausible wrong word is exactly what a cleanup pass leaves alone.
• It is roughly four times faster. 203 ms median against 890 ms for the multilingual model, on this corpus on 10 August 2026. Exact machine to be confirmed before launch. If your speech is entirely English, that speed is real and Bunso's multilingual model is the wrong tool for you — its own default English model is the closer comparison, and it is not the one measured here.
• It is already installed, free, and asks for nothing. No download, no model files, no Accessibility or Input Monitoring grants. For English-only dictation on a Mac it is a perfectly reasonable tool and this page is not an argument that it is bad software.

Where it loses, and why it is structural

Across the thirty clips, the Filipino content was gone from 28 of them. Not degraded — gone, replaced with English-shaped words that sound similar and mean nothing. "Bakit ang init ngayon?" came back as "Bakatam, Unique, and Paramvallayum".

This is not a setting you have failed to find. It is a model that does not contain the language. That also means it is not something a macOS update is likely to fix quietly — closing this gap needs a Filipino model that Apple does not currently ship in any locale.

All thirty, copied exactly

Nothing here is cherry-picked. Every clip that was recorded is in this table, including the ones Bunso got wrong: length for lint above, isi-same for i-cc, and a handful of hyphens and question marks that landed in the wrong place.

What this measurement is not

Thirty clips from one speaker is enough to show a difference this large and nowhere near enough to be a benchmark. It is one bilingual speaker's register, in one room, on one day. A different speaker, a heavier accent, more English per sentence, or a noisy café would all move these results and none of those have been tested.

We publish the transcripts rather than a score on purpose. Error rates on code-switched text depend on how each word was classified as English, Filipino or a hybrid of the two, and those classifications here were made by a non-native speaker and have not yet had a native review. The transcripts need no such interpretation: you can read both columns and judge them yourself.

More on what code-switching does to speech recognition, or the front page for what the app actually does with the text once it has heard it.