What changed, and what it measured

The engine now takes game events from the picture as well as the audio. A silent multi-kill that nobody reacted to out loud scores like the fight it was. Where a game is not supported the signal is simply absent and selection behaves exactly as before — nothing gets worse, some recordings get better.

When several people talk at once the engine plays it safe — it colours a line only when it is sure who spoke, so your name never lands on someone else's sentence. Every voice profile added makes it surer, and the hard overlaps keep getting better.

Speaker matching used to fail in both directions at once — missing the right person and occasionally colouring the wrong one. It now weighs a voice against the others actually present in the same recording, which removed both failures together.

Masking a slur does not save a clip that is about the joke. Such passages are now removed from selection entirely, before any clip is built — and the gate was hardened after one slipped past an earlier version.

Laughter is now heard in the audio itself instead of hoped for in the transcript, and reactions are measured as a build-up rather than line by line — a single exclamation carries no signal, a run of them does.

On silent passages the transcriber was inventing lines that nobody said — over two hundred of them reached the screen in a single stream. They are now caught and discarded before anything is written to your captions.

The transcriber was working from the wrong context, which does not degrade gracefully — it produces confident nonsense. The engine now works out what is actually being recorded before it commits to a word.

All heavy processing now runs on our own servers. Nothing touches the customer's machine, and capacity is rented by the minute when demand needs it.

The engine assumed one recording profile and silently produced nothing on anything else. It now reads the actual properties of every file it is given, so a phone capture and a studio stream both work.