| Who | Computer | App | Last seen | Takes |
|---|
| When | Who | Kind | Original | After the LLM | Audio |
|---|
Checking…
Beam = how many guesses the model keeps while it writes. More = a little more careful, slower. Live re-reads the speech every ~0.3 s, so a small beam keeps it from falling behind; Recording (one upload, FreeFlow / iPhone) reads once.
When idle — a model nobody used for this many minutes leaves the VRAM. RAM: back in about a second · Disk: back in a few seconds (frees the RAM too) · 0 = always in VRAM. The first words after that wait for it.
The final text goes through an LLM that fixes misheard words, spelling and punctuation — same language, same script, nothing added. A reply that changes more than the limit is thrown away and the original kept. Live = clients MAY use it for a live take's final text (once, when you let go — the live words while speaking stay as they are); each client turns it on for itself (vApp Voice: "Fix with LLM", off by default) · Recording = the one-shot upload (FreeFlow, iPhone) goes through it.