SpeechfyGabriel Anhaia

You held the key.
You said it.
It landed written.

Speechfy was a dictation app for the desktop, run entirely in Europe. It is retired now, and this page is the record of what it was and how it worked.

New message

Hold the key to try it. This is a re-enactment and uses no microphone.

Three moves and no window to switch to

  1. 1

    Hold the key

    Right Option by default, or any key you recorded. Speechfy checked your choice against what the system had already claimed.

  2. 2

    Talk

    A small pill showed your voice as a row of bars. It never took focus, so the app you were typing in kept its caret.

  3. 3

    Let go

    The text went in where your cursor was: the same message box, the same document, the same terminal.

From key up to finished sentence took 1.6 seconds at the median. The transcript streamed while you spoke, so only about 0.8 seconds of transcribing and 0.9 of polishing were left.

One sentence, written for where it was going

There were four registers: Off, Verbatim, Clean and Formal. Speechfy picked one by the app you were dictating into, so a terminal never got punctuation it had not asked for and an email never kept your ums.

Register: Formal

I pushed the fix, but it does not work on staging. It works fine locally. I will look again tomorrow.

What else was inside

One language in, another out
Two controls, Listen in and Write in. Speak Portuguese and the sentence arrived in English, with a note beside the switch saying translation was harder than tidying.
Auto language
It used the language you last dictated into the app in front of you, and the pill named it before you spoke. Listening in every language at once took about five times as long and still guessed wrong on short clips.
A dictionary of two halves
Names to get right, and corrections to apply whenever they came up. Both went to the model that wrote your sentence.
Snippets
A short spoken trigger became the longer thing: an address, a signature, a standup template.
Notes that were just files
Plain Markdown in a folder you chose, on your own machine. Speechfy never uploaded them and never deleted one.
History
The text of every dictation into another app, searchable and grouped by day. Notes stayed out of it, and so did anything said with an excluded app in front.
A receipt, not a scoreboard
Words, days used and time spent speaking. No streak and no goal.

Your voice stayed in Europe, and was not kept

The company, the servers and the models were all in the EU. Audio and the text made from it lived as long as the request took. Speechfy's policy was that none of it was read by a person or used to train a model.

  1. Your desktop

    A native client held the key, drew the pill and streamed the audio.

  2. Speechfy's service

    An EU service checked the account and counted the words.

  3. Speech model

    A model in the EU turned the audio into a transcript while you were still talking.

  4. Writing model

    A second model in the EU tidied it, using the register and your dictionary.

  5. Your cursor

    The sentence went into the app that already had focus.

How it was built

Mac
Swift, SwiftUI and AppKit. One background process with no Dock icon until a window was open.
Windows and Linux
Flutter clients carrying the same mechanic. On Linux the key was read from the input device, which is what let one build work on X11 and through XWayland.
Service
TypeScript on Hono, Postgres and Stripe. The client sent a transcript and the context around it; the service built the prompt and chose the model, so a register could be retuned without shipping an app update.
Security
TLS 1.3 in transit, Argon2id for passwords, SHA-256 for tokens.
  1. First commit, a Mac app with one job.

  2. Version 1.0 signed, notarised and downloadable, with 1,804 Swift tests and 384 service tests passing.

  3. Windows and Linux rebuilt as one Flutter app.

  4. Last commit to the site, by then in 14 languages. 854 commits across six repositories in all.

  5. Retired.