ElevenLabs Ships Voice Model v4

ElevenLabs released its new v4 voice model on September 28, along with a lighter v4 Turbo version. It supports more than 90 languages, up from 70 in the previous v3 version. The biggest quality jump shows up in Japanese, Brazilian Portuguese, Mandarin, and Cantonese. You can now clone a voice from a ten-second recording, instead of the longer sample the old version required.

For me, the drop in latency matters more than the extra languages. In voice agents, the user feels a pause before the reply within the first second of talking. With v4, audio generation starts as soon as the language model produces the first words of its answer, not after the whole text is ready. Cloning a voice from ten seconds of audio instead of a long recording session is practical too. With a short sample, it's much easier to spin up a voice for a support agent quickly, without making someone read out a long script.