ElevenLabs: Eleven v4

Eleven v4 is ElevenLabs’ expressive text-to-speech model for produced content, narration and character performance. Its new architecture uses the script’s context to shape who is speaking, what has happened and how the next line should sound. The input is a written script associated with a selected voice; the output is audio. The manufacturer highlights emotional dialogue, audiobook production, animation and games, and multilingual projects as characteristic applications.

Direction can be written into the script with inline audio tags. Tags such as [laughs], [whispers] and [door slams] can describe vocal delivery, a non-verbal event or a sound within the scene. V4’s documented update improves following a sequence of those directions compared with v3. For a short dramatic beat, write the actual spoken words and place a delivery cue at the point where it matters, rather than appending an unrelated list of emotions to the end.

Pauses and emphasis use natural-language audio directions. ElevenLabs documents [pause] and [long pause] for breaks and cues such as [excited] or [sighs] for delivery. SSML <break> tags are disabled for v4, so an older script using those tags needs to be adapted to the v4 direction format. A tag names the intended performance; it does not by itself set an exact number of milliseconds or guarantee a fixed final duration.

The producer’s dialogue workflow supports multiple speakers, allowing a scene to contain separate voices and sound events. Its v4 dialogue WebSocket profile supports up to ten registered voices, while v4 Turbo’s profile uses one. Prepare a clear order of speaker turns and keep each line associated with the intended voice. An illustrative exchange might move from a quietly uncertain question to a confident answer and then a relieved reaction. The content and emotional transition should be understandable from the script itself. For a single narrator, use the same clarity to distinguish narration from a quoted line or a change in mood.

Eleven v4 supports more than 90 languages, including English, French, German, Japanese, Spanish, Portuguese and Russian. The manufacturer describes consistent voice quality through language changes. Choose a voice and pronunciation appropriate to the language and region of the intended recording; the language of the script and the character of the voice both matter. For a localised lesson, translate the spoken wording first and preserve the meaning, terminology and delivery cues in the new script.

A generation supports up to 10,000 characters. Longer work is prepared as multiple connected generations, with context stitching intended to keep pacing and delivery consistent between segments. Divide a chapter at a natural sentence or scene boundary and retain the context around it. The manufacturer also describes stable speaker identity when a line is regenerated. That is useful for revising one passage while maintaining the same cast and performance style across the surrounding material.

V4 supports Instant and Professional Voice Clones in ElevenLabs’ producer workflow. ElevenLabs highlights their return in v4; its v3 prompting guide says Professional Voice Clones are not fully optimised for v3. Instant or Professional Voice Clones created before v4’s launch need to be retrained for v4 to work effectively with the model. Voice cloning creates or prepares the voice used for synthesis; it is distinct from the act of converting the supplied script into speech. No single default voice is required by the model’s identity, so select the voice for the narrator or character rather than assuming that the model name determines a speaker.

The manufacturer documents MP3 for general listening and media delivery, WAV or PCM for uncompressed production audio, and μ-law for telephony. Its API controls the sample rate and bitrate of the selected format. Pronunciation dictionaries can define how names, acronyms and technical terms are spoken. V4 also recognises standard IPA transcriptions enclosed in forward slashes, such as /IPA_transcription/, including stress markers for multi-syllable terms; the pronunciation can still vary with the voice and phrase. Prepare required terms before production and keep their preferred spoken forms consistent through revisions. The v4 producer profile uses Stability and Similarity voice settings; Style and Speed sliders are not available. Pace and emotional delivery are directed through the script and audio tags.

For a practical narration brief, supply a finished scene, the intended voice, the language, the emotional arc and the destination of the audio. For example: “Read the opening calmly, pause before the discovery and make the final line sound relieved; keep the narrator’s tone consistent.” Review pronunciation, speaker changes, audible tags or events and the transitions between passages. The useful deliverable is a produced speech recording tied to the script, with a selected take and the directions needed to revise a particular line.