Talk To It: dictate the brief — your voice never leaves your machine.
Beginners say “voice” in every thread, and they mean talking to the AI: dictating the prompt, the spec, the bug report. This entry vendors the free local engine for that — faster-whisper running OpenAI's Whisper model, both MIT — and proves it on a real recording before asking you to install anything.
Forge entry · Bundle Wave 1 · engine: faster-whisper (SYSTRAN, MIT) + Whisper weights (OpenAI, MIT) · proof run in the harness toolchain, 2026-10-23. One of the Forge shelf.
The worked proof: a recording, transcribed, graded
The test recording was this site's own spoken ad read — a generated voice (Kokoro, af_heart) reading a known script: 26.0 seconds of audio. The script gives us ground truth, so the transcription below is graded against it word for word, not vibes. Run on a small 2-core CPU server — no GPU, no cloud.
| Speed | PASS — 3.9 s to transcribe 26.0 s of audio (~6.7× real time), plus 2.5 s model load. English detected at 100% confidence. |
| Accuracy | PASS, with the error printed — one content word wrong in ~78 (below, in context). One formatting difference (“$5” for “five dollars”) — same meaning, not counted an error. |
| Privacy | PASS — the engine is local. The recording was transcribed by the machine it sat on; nothing was uploaded anywhere. |
What the engine heard (verbatim)
“Your AI runs out of credits in the middle of your project. What if it could finish anyway? Start by answering a few straight questions. Your AI turns the answers into a plan, checks its own work against them, and fixes what fails. Another AI can pick up right where the last one stopped without paying the build at all. Spend $5 once or subscribe for the month if you actually like it.”
The one wrong word, in the open: the script said “without paying the bill at all”; the engine heard “build.” Two real words, one plausible sentence — exactly the class of error to skim for after dictating (names and load-bearing nouns). One word in a paragraph, fixed in seconds by eye, is the honest price of free and local — and this page prints it rather than rounding it away.
The flow (what you actually do)
- Open the clarifier first — the /start questions, or the game clarifier in the Godot pack. Dictation answers questions better than it free-associates; the questions are the structure.
- Record on your phone, one answer or one long take, conversation pace.
- Transcribe locally with the pack's exact install (Python venv, faster-whisper, a short script — the same commands that produced the proof above).
- Skim once, fix the names — errors cluster on project names, people, jargon: the words context can't guess.
- Paste where the typing would have gone. Brief, spec, bug report — dictated in five minutes instead of typed in forty.
The nits, stated plainly
- One wrong word in ~78 is the base model's real level on clean speech. Noisy rooms, accents, and crosstalk will do worse. The small model is the built-in step up (same commands, one word different in the script) — slower, fewer errors.
- This site does not transcribe for you. There is no upload box here, deliberately: a server doing this for every visitor runs a per-visitor compute bill, and your voice notes are yours. The pack sets the engine up on your machine.
- A client-side follow-on exists and is scoped, not built: Whisper can run entirely in the browser (WASM/WebGPU builds), which would keep the same privacy shape with no install. It is noted here as the entry's next version — not promised, not dated, and not quietly implied to exist today.
- The install has one sharp edge — a version pin (av==13.1.0) because the newest audio library changed an API the engine uses. It's documented in the setup file with the reason, not cargo-culted.
Download the pack
| File | What it is | Size |
|---|---|---|
| README.md | What it is, the engine + licenses, the proof in brief | 2,155 B |
| setup-on-your-machine.md | The literal install: venv, faster-whisper 1.2.1, the av pin, the transcribe script | 1,786 B |
| dictate-a-prompt.md | The beginner workflow: answer the clarifier out loud, skim-fix the names, paste | 2,150 B |
| proof-transcript.txt | This page's worked proof as a downloadable text file | 1,133 B |
⬇ Download all four — talk-to-it-pack.zip🚀 Then dictate your answers at /start
Adding to the Forge
A Forge entry needs a name, a one-line claim, an author line, a worked example with real measurements, the files, and its limitations on its page — this one included.