ElevenLabs released Eleven v4, a new text-to-speech model, on 28 September, together with a faster v4 Turbo variant, Martin Cid Magazine reports. The model can clone a voice from about 10 seconds of audio and have it speak more than 90 languages, up from 70, while keeping the speaker's timbre; the company named Japanese, Brazilian Portuguese, Mandarin and Cantonese as its biggest gains. Free accounts can use it.
The main change is direction. Writers can now put performance cues in square brackets, such as a laugh, an angry line in a French accent or background sounds like rain, and the model also infers tone from the surrounding text. ElevenLabs says scenes with several speakers now hold each voice steady over long passages, a known weakness of synthetic narration. v4 Turbo, aimed at voice assistants and call centres, starts speaking in roughly 150 milliseconds.
For creative work, the magazine sees podcasters dubbing episodes in their own voice, small game studios voicing minor characters without booking a studio, and audiobook narrators scripting pauses and chuckles. Benchmarking firm Artificial Analysis put v4 at the top of its voice leaderboard within hours; the company's claim that about three in four listeners preferred it in blind comparisons comes from its own testing.
The short sample is also the risk. ElevenLabs says clones need the voice owner's consent, runs a public classifier for audio made with its tools and blocks some political figures, but those checks depend on uploaders telling the truth. The free tier allows around 10 minutes of audio a month, with paid plans from $6. The London company raised $500 million at an $11 billion valuation in February and, per TechCrunch, has passed $600 million in annualised revenue.
Sources · 1