INFINIGHTMake your first book, free

Ninety seconds of a grandparent's voice

What actually happens when you record a family voice, why we ask for consent on the record, and what we will not do with it.

One of the quieter features in Infinight is also the one people react to most. You record somebody reading for about ninety seconds, and after that the app can read any book in their voice.

For families spread across a few countries, that lands differently than a feature list suggests. A grandmother in Kraków reads the bedtime story in Rotterdam, on a Tuesday, without anybody having to be awake at the same time.

Here is exactly how it works, including the parts we say no to.

What the recording is

Three short passages, read into a phone. The app watches the length and the level while you go, because a recording that is too short or too quiet produces a voice that sounds like a bad phone line, and it is better to say so during the recording than after it.

The passages exist to get enough natural speech, not to teach the model the words. The technology takes audio only. There is no script it is matching against, which is why you can read anything as long as you sound like yourself doing it.

Before a voice is created we record who it belongs to and that they agreed. That is stored on the voice itself, not in a log somewhere.

This is not us being careful for the sake of it. A cloned voice is biometric data under European law, in the same category as a fingerprint. Someone else uploading a recording of you is exactly the failure this is built to prevent.

Deleting a voice removes it from us, from the company that hosts it, and from anywhere it had been set as the default narrator. If it was the voice your books were being read in, the books go back to a standard narrator rather than quietly failing.

What we will not do

A voice never moves between families. If the same grandparent belongs to two households, each one records separately. Copying the voice across would be processing it on a permission that was never given, and “but it is the same person” is not a defence.

A voice never travels with a published book either. If you publish a story to the public library, the narration stays home. That consent was given for one family’s bedtimes, not for the internet.

The part that is imperfect

A cloned voice reading a language its owner does not speak sounds like them speaking it with an accent. Some families find that delightful and some find it uncanny. You can listen before you commit to anything, and you can retune the delivery afterwards with four sliders: warmer, slower, more expressive, closer to the original.

It is not the same as having them there. It is closer than a stranger reading it, which is what the alternative was.

Back to the journal