Eleven v4 for expressive audio

Eleven v4 is aimed at audio that needs performance, not merely intelligible words. ElevenLabs says its new architecture reads a script with awareness of who is speaking and how a line should land. Writers can put directions such as whispers, pauses, laughter and sound events into the text, then cast more than one speaker. The release also emphasizes speaker stability when a line is regenerated, a practical concern for long projects where one changed sentence should not sound like a different person.

Long-form creators can use context stitching to keep delivery consistent across successive generations, while professional voice clones return as a supported option in v4. The product page says one generation accepts up to 10,000 characters and lists MP3, WAV/PCM and telephony-oriented formats. These details make the model plausible for narration, multilingual education, character dialogue and ads. All generated audio still needs editorial listening: pronunciation, pacing, emotional fit and possible sound-tag artifacts can vary, especially in specialized vocabulary or longer sequences.

The related v4 Turbo model is built for live conversations and agent loops. ElevenLabs reports roughly 100 ms median inference and about 150 ms to first speech, plus bidirectional streaming and voice continuity. Those numbers are vendor measurements, not a guarantee for a particular network, voice or application. Choosing the production v4 model versus Turbo should depend on the use case: a finished audiobook can tolerate extra processing for quality, whereas an interactive agent must avoid unnatural pauses. Voice cloning has a separate consent obligation; ElevenLabs says every clone requires verified consent from its owner. Commercial output rights also depend on plan terms, so free-plan testing should not be treated as a commercial license.