On-device transcription with speaker labels: the speech engine inside Transcription Studio is now FCT's own

You recorded the meeting. Four people were in it. What comes back is one long block of text with no names in it, and the one line you actually needed, the part where somebody committed to a date, is in there somewhere without a face attached.

That is the ordinary result. The usual fix is to send the audio to a company’s servers and wait. Transcription Studio does it on your device instead, and it marks who spoke.

We call it Speaker Labels. Every line of the transcript carries the person who said it, from the moment the recording stops.

What it gets right

Those numbers were measured on September 11, 2026, on a Mac mini M4 with nothing else running, against VoxConverse: a public set of real recordings where researchers have marked by hand who is speaking and when.

The number we care about most is the middle one. People talk over each other. In sixteen real recordings there were 132 seconds where two voices genuinely overlapped, and Speaker Labels marked both of them in 87 cases out of a hundred. The engine the app ran on before found none of that overlap on the same recordings. It never once marked two people speaking at the same time, which is the difference between a transcript of a conversation and a transcript of a speech.

A transcript without it: One voice

• Every line reads as the same person
• Interruptions vanish into the sentence before them
• You scrub the audio to find who said it
• Audio goes to a server to be read

Speaker Labels: Four voices

• Each line carries the person who said it
• Two people talking at once are marked as two
• You search the name and read the line
• The audio is read on your own device

How it works

Three models do the work, and FCT converted all three from the published research checkpoints and wrote the code that drives them. They run on the Neural Engine, the part of an Apple chip built for exactly this.

1. Your recording: Mic, meeting audio, a file or a link.
2. Language picks the model: Not a guess from the sound.
3. Words: On the Neural Engine, 124 times faster than real time.
4. Voices: Four speakers, overlap included.
5. A transcript with names.

Choosing the model from your language rather than from the audio matters more than it sounds. A recognizer built for European languages will happily produce confident nonsense when it is handed Japanese, because nothing in it can tell. Transcription Studio never asks it to: the language you work in selects the model before the first sample is read.

Speed comes out of the same decision. An hour of audio transcribes in about thirty seconds, and the speaker pass over that hour takes under two minutes, both on a Mac mini M4 with a quiet machine.

Where the work happens, and where the result is kept

Those are two facts and we never collapse them into one. Speech recognition and speaker identification run on your device, so the audio is never sent anywhere to be read. The recording and its transcript are then stored in your own private FCT account, which is how they reach your other devices.

Availability

Transcription Studio is built and is FCT’s own daily transcription tool. It is not on the App Store, and we are not naming a date we have not earned. The three converted models it runs on are public now, at fcttechnologies/fctspeech-coreml on Hugging Face.

Owning the engine is what makes the rest of this possible. When the speech stack is someone else’s, the language routing, the overlap handling and the promise about where your audio goes are all decisions somebody else already made for you. We would rather make them, and then measure them.