# Talk To It — speech-to-text for people who'd rather talk

You think faster than you type. The beginners in our market survey said
"voice" in almost every thread — and they meant *talking to the AI*:
dictating the prompt, the spec, the bug report, instead of typing it.
This Forge entry vendors the free engine for that and proves it works,
so you can dictate your next build brief tonight.

## The engine

**faster-whisper** (MIT) running OpenAI's **Whisper** model (MIT — the
weights carry the same license as the code; weigh that, because in this
space models are often licensed tighter than their code). It runs
**locally, on your machine, on CPU** — your voice never leaves your
computer, there is no per-minute meter, and there is no account.

The `base` model is the starter: a ~140 MB download, fast enough on an
ordinary machine. `setup-on-your-machine.md` is the exact install we
ran (same commands, same versions), and `dictate-a-prompt.md` is the
workflow once it works.

## The proof (measured here, printed in full)

We recorded a real paragraph — a spoken ad read, 26.0 seconds, a
generated voice reading known text — and transcribed it with exactly
the setup in this pack, on a small 2-core CPU server:

- Transcription took **3.9 s** (~6.7× faster than real time), model
  load 2.5 s, English detected at 100% confidence.
- Against the known script: **one content word wrong in ~78** — it
  heard "…pay the build at all" where the script said "…pay the bill
  at all." One substitution formatted differently ("$5" for
  "five dollars", same meaning). Everything else word-exact.
- Verdict, plainly: good enough to dictate a brief or a spec and fix
  one word in a paragraph by hand. Not good enough to transcribe
  legal testimony unattended. The `small` model is the step up if
  your machine has the patience — same commands, one word different.

## Where the audio goes

Nowhere. That's the point of a local engine: the file you record is
transcribed by your own CPU. This site offers no upload-and-transcribe
service, on purpose — a metered server would cost money per visitor,
and your voice notes are yours.