Spanish voice dictation that needs no fixing afterwards

Dictating in Spanish half works in almost every app, and the reason is boring: transcription is not the problem. Speech engines have understood Spanish well for years. What goes wrong is everything that comes after.

The four places it breaks

The opening marks

In Spanish a question opens with ¿ and closes with ?. Almost no dictation app puts them in, because general purpose models are trained on English and English does not have them. So you get "cómo estás?" and you fix it by hand, every day, forever.

Getting them right is not cosmetic: you have to know where the question starts, which is not always the start of the sentence. In "Hola Sofi, cómo venís con el informe" it opens at "cómo", not at "Hola".

Voseo

"Acordate", "creá", "mandáselo", "reservá". An engine that assumes the tuteo forms hands you back "acuérdate", or worse, eats the accent. It is the kind of error that breaks nothing and forces you to reread every sentence.

Filler words

Talking is not writing. "Che, o sea, me parece que el partido de mañana, viste, lo tendríamos que mover" is a perfectly ordinary spoken sentence and an unreadable written one. Removing the fillers without touching the rest is harder than it looks, because "che" and "viste" are sometimes fillers and sometimes not.

Work Spanglish

If you work in tech, your Spanish has English inside it and you conjugate it: mergear, pushear, deployar, un pull request, un branch. An engine that hears "creá una brunch nueva" and writes "brunch" did not get the language wrong, it got the domain wrong. The page for developers goes through what that vocabulary looks like in a day of prompts and commits.

What DeftNib does about it

It treats all of that as the main case, not as a localization. Opening marks, voseo, fillers and mixed vocabulary live inside the pipeline, and there is a file of cases that a quality evaluation runs against every time the code is touched. The pairs on this page come out of that file, not out of a demo assembled for the page.

The pipeline is the layer above the engine: the same Apple engine that powers the dictation built into macOS, with fifteen rule engines and a local model on top. When you enumerate things without saying the word list, it comes out as bullets. When you dictate an address, it comes out as an address.

And all of it happens inside your Mac. There is no account and no subscription, and your audio and your text are never transmitted. What leaves your Mac, and what never does.

The other languages

The full pipeline runs in English and Spanish. In the other languages your Mac supports you get Apple's native transcription, which already punctuates and capitalizes well in its own languages, plus your dictionary and your snippets. We say so plainly because the cleanup engine is calibrated on English and Spanish vocabulary, and pointing it at French would do subtle damage instead of good.

Questions