Choosing a synthetic voice that sounds real

A synthetic voice turns text into speech. The best ones are now so natural you could mistake them for a host. But they are not all equal, and the choice comes down to a few simple criteria.

Two families of voices

• Ready-made voices: dozens of voices offered by the tool, free to use. You pick a tone, an age, an accent.
• Cloned voices: the tool learns a voice from a recording (one to a few minutes is enough). Ideal for keeping your own voice across all your content.

What to listen for

• Naturalness: breathing, changes of pace, the rise of a question.
• Language and accent: a clone keeps the accent of the recording. A voice learned from a French speaker talking English will sound French in English: for flawless English, a native voice is better.
• Pace: some voices speak too fast for listening. Around 15 to 17 characters per second is easy to follow.

The right method: have the same ten-second excerpt read by three or four voices, and listen to them back to back. The difference is obvious to the ear, far more than on a spec sheet.

The red line

Never clone a real person's voice without their written consent, even a celebrity, even "as a joke": it infringes their rights and can end up in court. And when a synthetic voice imitates a real person, the European AI Act requires you to disclose that the content was generated.