How a voice imitation is scored

A voice imitation game asks one thing: hear a sound, then make it. A goat, a car alarm, somebody yelling across a car park — everybody has a go, and a number comes back saying how close each of them got. That number is the part people argue about. Read the reviews of almost any game in the genre and the same complaints come up: the score feels arbitrary, the game does not seem to hear you, and anybody with a very deep or very high voice suspects it is marking their voice rather than their attempt.

Those are design problems, and they have design answers. This is how Mimic, PartyDub's imitation mode, turns a two-second attempt into a score — written down so you can check whether it is judging what you think it should.

It measures rather than guesses

No language model and no speech recognition are involved at any point. Your recording is compared with the original by measurement, and the result is one number out of 100, ready the moment the recording window shuts. Nothing is trying to recognise words, so there is nothing for an accent to confuse. The clip is cut into pieces wherever a new sound starts, so three barks are judged as three barks, and each piece is compared on four things:

• —Timing — whether each of your attacks lands where the original's does.
• —Pitch movement — whether your voice rises and falls where the original rises and falls.
• —Loudness shape — whether you swell, cut off and punch in the same places.
• —Rhythm — whether each piece lasts as long as it should. Rushing and dragging are equally wrong, so holding a sound twice as long costs the same as cutting it in half.

Your voice is not what gets marked

Before two pitch lines are compared, each voice's own typical pitch is taken out, so what is left is the shape of the movement rather than the notes. A bass and a child doing the same rise and fall score the same. Loudness is handled the same way — it is compared as a shape, not a level — so a quiet microphone is not a worse score and shouting is not a better one.

One more rule, and it is the one that keeps the game fun: going bigger never costs you. A take that follows the original's movement but exaggerates it scores in full. A take that follows it timidly, at half the size, does not. Being flat is the mistake. Overdoing it never is.

Not every sound has a voice in it

A siren, a sung hook, a door slam and a sheep ask for different things, so the score shifts its weight depending on what the clip actually contains. A melody leans on pitch movement more than anything else. A noise with no pitch in it at all — static, a crash — ignores pitch entirely and is judged on loudness shape, timing and texture instead. Anything in between is weighed evenly. Whatever the clip, the total is out of the same 100, so the clip you are dealt does not decide how high you can score.

Late is not the same as wrong

Before anything is measured, your whole take is lined up with the clip. Coming in a little late across the board costs very little, because it is barely audible — and the results screen says so in plain words: you were consistently about this many milliseconds behind, come in sooner. Being scattered is different. Early on one attack and late on the next is what actually sounds wrong, so that is where the points go. When two players tie, the one whose attacks were more consistent places higher.

What a zero means

Zero is not the bottom of the arithmetic. It means the take did no better than an unrelated sound would have. Any noise of roughly the right shape lines up with the original a little by accident, so the scale starts where chance ends: everything above zero is the part of the clip you genuinely matched.

Before a take is scored at all, it is checked for a performance. A quiet room is not silent — a fan or a fridge is a steady tone, which is exactly what a pitch tracker is built to find — so a take has to rise clearly above the room's own noise, move, and start at least one sound. One that does not is not given a bad score. It is not scored, and you are told why — that the room drowned you out, for instance.

Playing the clip back into your microphone gets noticed. A take that sounds too much like the original is flagged on the results screen for the whole room to see. The score stands, because a genuinely good impressionist can look the same, and accusing the best player at the table is worse than a party game being gamed now and then.

How to score higher

• —Listen properly the first time. You hear the clip once, then everybody records.
• —Nail the first attack. A clean start keeps everything after it in line.
• —Copy the movement, not the voice. Rise where it rises and cut off where it cuts off, at your own pitch.
• —Go bigger, never smaller. Exaggeration is free; a timid take loses points.
• —Use wired headphones. Bluetooth adds a delay you cannot feel, and speakers leak into your take.

None of this makes the game any less silly. It only means that when the room laughs at your goat, the number under it is one you can argue with.