EN 
SV 01 Audio · machine learning
2019It's been possible to clone voices for a while now
In 2019, I ran a small experiment — and very few people got everything right.
Update · 2025
Note from 2025
Since this experiment, voice cloning has become completely commoditized. Services like ElevenLabs and Play.ht now offer near-perfect voice cloning from seconds of audio, available to anyone. What was a research project in 2019 is now a feature in consumer apps. The question is no longer "can we clone voices?" — it's "how do we deal with a world where any voice can be faked?"
Inline voice test
1 / 12
The experiment
I cloned several voices using neural speech synthesis and speaker embeddings — from just 5 seconds of audio per person. Then I asked 108 people to listen to 12 clips and guess which were real and which were AI-generated.
Most people were confident in their answers — but the results told a different story. On average, participants were fooled by about half of the cloned voices. The technology was already good enough in 2019 to deceive most listeners.