Quality and measurement3 min read

Speaker similarity

Also known as: SIM

In short

Speaker similarity measures how close a generated voice is to the real speaker, usually as a score between voiceprints. It answers whether a clone still sounds like you, separate from whether the words are right or the delivery is natural.

Why it matters

Similarity is the metric that answers the first question anyone asks of a clone: does it still sound like me. It is worth separating from accuracy and naturalness, because a voice can be perfectly clear and natural and still not sound like the person it was cloned from.

It also needs a human baseline to read correctly. Two recordings of the same person on two different days do not score a perfect match, so a good clone is one that lands inside the range of your own voice, not one that hits an impossible number.

In practice

Compare a generated sample against a held-out reference of the real speaker, not against your memory. If a tool reports a similarity score, ask what a same-speaker baseline scores, so you know what good looks like.

How Vocast handles this

Vocast reports speaker similarity on every render. Two recordings of the same real person score about 0.909 against each other, which is the range a good clone sits inside.

Related terms

Your voice, on your Mac

Vocast clones your voice from about ninety seconds and narrates any script in it, fully on-device, for $49 one time.