Good to know
Which model? Qwen3-TTS (1.7 billion parameters) from Alibaba's Qwen team, freely available under the Apache 2.0 license. It runs on our own NVIDIA Tesla P40. We generated the ready-made voices once from a description using the VoiceDesign model.
How good is the voice clone? Just 5–10 seconds are enough for a surprisingly similar voice. Results are best with a clean recording without reverb or background noise, and when you enter the exact wording of the sample.
What happens to my data? Your text and voice sample are sent encrypted to our server and from there to our graphics card in Germany. The voice sample is deleted right after you speak, and the finished audio after one hour (or immediately if you click "Delete now"). None of it is public. Only with "Describe a voice" does Claude (Anthropic) translate your voice description into English, because the voice model only understands English descriptions. Your text stays with us.
What's not allowed? Cloning other people's voices without permission, using them to deceive or defraud others (for example, fake calls or voice messages), insults, and illegal content. That's not just unfair, it's often a criminal offense.