The Interview

What this is

You are a night-shift officer at a border station. A candidate sits across the desk with a dossier of eight statements about her life, and your job is to work out which ones hold up. She answers whatever you actually ask — by voice or by typing — and none of it was written in advance.

Everything runs on your computer. No account, no API key, no internet after the first launch, nothing sent anywhere. This is a demo of an offline AI dialogue system for game developers; the link to it is at the bottom.

Before you download

Download	788 MB zip (Windows 64-bit), 1334 MB unpacked
First launch also downloads	1749 MB language model, once, into your user data folder
Disk space	About 4 GB free while it installs
Platform	Windows only for now. Not macOS, not Linux, not browser
Hardware	A GPU with 4 GB of video memory or more is comfortable. Less still runs — see below
Internet	Only for the first launch, to fetch the model. After that it is fully offline

The download is the same on every machine — the small model — and that is deliberate. Every gigabyte between "downloaded" and "heard her speak" is somebody deciding not to bother, and the small model holds the character just as well; the difference is the download, not the writing. What the game does measure is what your machine can run, and it tells you in a sentence: whether you have room to raise the model in Settings, or whether you are below the line. A machine that cannot host a model at all still gets a finished interview — she speaks written answers instead of generated ones, and the game says so rather than pretending or hanging.

About the model download

The first launch fetches a language model from Hugging Face — Qwen3-1.7B, published by the Qwen team under Apache 2.0. The download shows its size before it starts, checks the free space first, resumes if your connection drops, and verifies the file when it finishes.

If you want a larger one — Qwen3-4B (2.3 GB) or Qwen3-8B (4.8 GB) — choose it in Settings before you begin. They are slightly more characterful and a great deal more download, and the 8B needs more free space than the 4 GB quoted above.

Your first launch will show a Windows warning, and here is why

The executable is not code-signed, so SmartScreen will say "Windows protected your PC" and name an unknown publisher. That is what it says about every application from a developer who has not bought a certificate, which for a free demo is not a cost worth passing on to you as a price.

Click More info → Run anyway if you are comfortable; skip it if you are not, and nothing is lost.

The game also downloads a large binary on first run, which is exactly the shape of thing antivirus software dislikes. It is the language model, from the address above, and nothing else reaches the network at any point.

The microphone

Push-to-talk on V. If the microphone is unavailable the game says which of the three things happened — no device, no permission, or a device that is not passing any signal — and you can type instead. It never falls back silently.

Speech recognition runs locally too, which is the slowest part of the loop: expect several seconds after you release the key before she starts answering. Typing skips that wait entirely.

Controls

Key	What it does
Type + Enter	Ask her a question
Hold V	Ask it out loud instead
Esc	Pause — resume, return to the menu, or quit

Everything else is clicked: CONFIRM or CONTRADICT on each line of the clearance form, then SIGN: CLEAR or SIGN: DETAIN. The form scrolls; the last line is below the fold.

What it demonstrates

• She answers what you actually asked. Nothing is chosen from a list of authored replies.
• She keeps her story straight. The eight statements are hers; she does not contradict them, and pressing her on one does not make her invent a different past.
• She notices the room. Ask about the recorder on the desk and she knows it is there — because it is, not because a writer guessed you would ask.
• You decide. Sign the form as CLEAR or DETAIN. There is no correct answer waiting to be found.
• Play it twice and she answers differently. Same question, new interview: "Yes, the reel-to-reel tape recorder is running. I don't trust anyone here." and "Yes, the reel-to-reel tape recorder is running. I've been listening." Same facts, said by someone in a different mood.

Credits and licences

Made by Tim "merrymaker14" Musharapov.

The demo ships several open components. Each is named on the CREDITS screen in the main menu, and the full notices sit inside the build under InterviewDemo_Data/StreamingAssets/LICENSES:

• Piper and piper-phonemize (MIT), ONNX Runtime (MIT) — speech synthesis
• espeak-ng (GPL-3.0) — text-to-phoneme, used by Piper as a separate process
• whisper.cpp / whisper.unity (MIT) — speech recognition
• LLM for Unity (Apache-2.0) — the local model runtime
• Voice en_US-ljspeech-medium — public domain, no attribution required
• Qwen3 language model (Apache-2.0)

Source for the GPL-3.0 component. espeak-ng 1.52.0 as shipped here is built from https://github.com/rhasspy/espeak-ng at commit 0f65aa301e0d6bae5e172cc74197d32a6182200f — the fork piper-phonemize compiles against, not upstream. The binaries are redistributed unmodified. A copy of GPL-3.0 travels inside the build.

Telling me it went wrong

There is no telemetry in this demo — nothing reports back, by design — so the only way I learn anything is if you tell me.

Leave a comment on this page, or write to tmusharapov@bk.ru.

The game can produce a diagnostics summary — your hardware, the tier it chose, the model and the version numbers, no personal data, sent nowhere — which you can copy into a report. It is far more useful than "it didn't work".

I answer asynchronously and from an unhelpful time zone: expect a reply within a few days rather than the same hour. Silence means I have not read it yet, not that it was ignored.

The plugin

This demo is built on Offline AI NPC, a Unity plugin for exactly this: local language model, local speech recognition, local speech synthesis, per-character memory and state. No cloud, no API keys, no per-token cost.

https://merrymaker14.github.io/offline-ai-npc/ — what it is, what it measures, and the full documentation.

On the Unity Asset Store: https://assetstore.unity.com/packages/slug/400726 — the page goes live when Unity's review completes. The address is already final, so the link above will not change.