Windows desktop app
Practice coding interviews with a second screen that keeps up.
Whisper-Line listens while you work through a problem. It transcribes what is said in real time, catches the question the moment it is asked, and streams the answer to a second screen — a monitor, a tablet, a phone in a stand. Your main screen never changes.
The moment it happens
You do not stop, alt-tab, and type out your problem to something. Whisper-Line is already listening. A question leaves your mouth and it is detected, sent and answered — no hotkey, no prompt box, no gap in your thinking.
And it knows the difference between a question for the assistant and you muttering at your own code. Talking yourself through a fix stays private.
You: okay so if I sort it first that's n log n, then…
You: what would the two-pointer version cost here?
Two screens, one session
Answers stream in word by word, the way a person thinks aloud — first words on screen in well under a second on a warm local model. You start reading before the model has finished writing.
The display is a separate page on your own network. Put it on the monitor to your left, a tablet on the desk, or a phone in a stand. Your workspace stays your workspace.
“…and what's the space complexity of that?”
Your code, your pace
Point Whisper-Line at the file you are working in. Every time you save, it has your latest code next to the problem statement — so a hint lands on the loop you actually wrote, using the variable names you actually chose.
Turn on screen capture and it reads the problem straight off your monitors using Windows' own offline OCR. No pasting, no browser extension, any judge site. It is off by default and never switches itself back on.
def two_sum(nums, target): seen = {} for i, n in enumerate(nums): if target - n in seen: return [seen[target - n], i] seen[n] = i
seen[n] = i does on a repeat.
Prompt transparency
Every prompt Whisper-Line sends the model is visible in the Prompt Builder, and editable. Not a summary of it — the literal outgoing request, system and user messages separated, with character and token counts, refreshed as you speak.
There is no hardcoded system prompt out of reach. Each one arrives prefilled with its built-in text, independently editable and independently resettable. Rewrite one and it applies on the next request. Turn the system message off entirely and see what the model does unprompted.
system You are a coding-interview practice assistant. Answer briefly, out loud… user ## CURRENT QUESTION (last 30s) what's the space complexity of that?
4 812 chars · ≈1 190 tokens · refreshed 5s ago
Features
Eight things happening in the background of a practice session, none of them asking for your attention.
Words appear as you speak, with no chunk boundary cutting one in half. It keeps up in real time on an ordinary laptop CPU — no GPU needed. Four engines to pick from; the fastest measured, Parakeet, runs about 13× faster than you can speak.
Knows a question aimed at the assistant from a directive like “walk me through it”, and both from you muttering at your own screen.
Replies land word by word as the model writes them, in every mode — instead of nothing, then everything at once.
A separate page for another monitor, a tablet or a phone on your network — and you choose what appears on it.
A live watcher on your source file. Save, and the next hint already knows what you just changed.
Every prompt visible and editable, with a live preview of the literal request. No hidden prompts anywhere in the app.
Windows' own offline OCR reads the problem off your monitors. Optional, off by default, and re-confirmed every session.
Ollama on this PC or a GPU box on your LAN, or a hosted model. Switch providers at runtime, no restart.
How it works
A per-user Windows installer — no admin prompt, nothing to configure. It lands in your own app folder with a Start Menu entry, and your settings and models live outside it so an update never touches them.
Create your account on first launch, with email or Google. The trial is the full app: every feature, nothing held back, nothing to enter but an email address.
Paste a problem or let it read your screen, point it at your code file, open the display on the monitor beside you — then talk, and work. It keeps up.
Transparency
Point Whisper-Line at Ollama — running on this PC, or on a GPU box elsewhere on your network — and nothing about your session crosses the internet. No prompt, no source file, no transcript.
Speech is transcribed locally on every provider, cloud ones included: your audio is never uploaded. And because every prompt is visible before it is sent, you can check exactly what would leave, rather than take our word for it.
Whisper-Line is a rehearsal tool. Mock interviews with people who know it is running, solo practice against a timer, code-review drills, getting fluent in a language you are new to.
It does not hide windows, defeat proctoring, or execute your code — none of that is in scope, and none of it is planned. Each session asks you to confirm what you are using it for, and that answer is never remembered for you.
This website sets no cookies and loads no analytics. The app itself keeps its logs, settings and downloaded speech models in a folder on your own machine. Read the privacy draft.
Free trial on first launch. Windows 10 and 11. Bring your own model, or use ours.