ElevenLabs Eleven v4 Text to Speech

ElevenLabs Eleven v4 Text to Speech

ElevenLabs Eleven v4 is the newest and most expressive model in ElevenLabs' text to speech API, turning written scripts into directed, lifelike audio. It is the first of a new model generation and a net upgrade over Eleven v3 in quality, voice accuracy, emotion, and language coverage.

What is ElevenLabs Eleven v4 Text to Speech?

Eleven v4 is a text-to-speech model that reads tone and pacing straight from your script and lets you direct delivery with inline audio tags. Write cues like [whispers], [laughs], [sighs], or [shouting] in square brackets and the model performs them instead of reading them aloud, following tag sequences and emotion shifts more reliably than earlier models. It speaks 90+ languages, supports inline IPA pronunciation, and ships in two variants: eleven_v4 for the highest quality and eleven_v4_turbo for low-latency, real-time use.

Key Features

• •Inline audio tags for emotion, delivery, pacing, and ambient sound, performed as direction, not spoken.
• •90+ languages with native accent handling — one voice speaks any supported language and adopts a native accent in it.
• •Inline IPA pronunciation between slashes to lock how names and technical terms are said.
• •Context stitching for consistent voice and pacing across scripts up to 10,000 characters.
• •Stability and Similarity controls, plus multiple output formats (MP3, Opus, WAV).

Best Use Cases

Eleven v4 is built for directed, expressive speech: character voiceover for games and animation, audiobooks and long-form narration, dubbing and multilingual localization, trailers and ads, and conversational voice agents on the Turbo tier. In testing, a single tagged script moved cleanly from a booming announcement to a confiding whisper to an amused aside, and the same English voice read fluent Spanish with native phrasing — making one voice reusable across markets.

Prompt Tips and Output Quality

Place one audio tag per clause; its emotion carries forward until the next tag. Give the model a full scene rather than isolated sentences, since v4 performs better with context. Use language_code to lock a non-English output language, and inline IPA in slashes to fix tricky names. If a line misses, swap the tag ([let down] for [despair]) and regenerate rather than rewriting the text. Output is clean, broadcast-quality audio at your chosen format.

How to Use ElevenLabs Eleven v4 Text to Speech

Eleven v4 turns a written script into a directed voice performance. The craft is in the script: write what you want said, then steer how it is said with inline audio tags and punctuation. You only need two fields to start, prompt and voice_id, and the rest fine-tune delivery.

Core workflow

Put your text in prompt and pick a voice_id (a premade voice, a Voice Library voice, or one you cloned or designed). Drop audio tags in square brackets where delivery should change, for example [warmly], [curious], or [whispers]; use one tag per clause and let the emotion carry forward to the next tag. Ellipses slow a line, dashes cut it off, and exclamation marks add intensity. Give the model a full scene rather than one isolated line, since it reads context to perform well.

Recommended Settings

• •model_id: eleven_v4 for finished content and audiobooks; eleven_v4_turbo for real-time agents and high-volume drafts.
• •stability: 0.3 to 0.5 for expressive, performed reads; 0.7 or higher for steady, uniform narration.
• •similarity_boost: keep near 0.75; raise for a tighter match to the source voice.
• •language_code: set the ISO code (for example es, hi, fr) for non-English output so the voice adopts a native accent.
• •output_format: mp3_44100_128 for web playback, wav_44100 when you will edit the audio.

Example voices to try

Each example below was generated end to end at the settings above; the audio tags were performed, not read aloud, and the scripts kept their numbers and proper nouns.

• •Audio-tag emotion shifts (default). Prompt: "[booming] Ladies and gentlemen, welcome — welcome one and all — to the Carrington Brothers Travelling Circus! [whispers] Now, lean in close... [curious] Can a single acrobat truly cross forty feet of wire with no net below? [laughs] Our Madame Cheng says yes... [softly] So hush now." The read moves grand to confiding to amused in one take. Tip: place one tag per clause and let the emotion carry forward.
• •Multilingual with native accent. Prompt: "[warmly] Bienvenidos a la Sagrada Família, la obra maestra de Antoni Gaudí... ¿Saben cuántos años lleva en construcción? [pause] Desde mil ochocientos ochenta y dos..." The same English voice reads fluent Spanish. Tip: set language_code to es to lock the language; keep tags in English even inside a non-English script.
• •Inline IPA pronunciation. Prompt: "[hushed] Deep in the lakebeds of Xochimilco lives... the axolotl. Say it with me: /ˈæksəlɒtl/... Biologists call the trait neoteny — /niˈɒtəni/." Hard names land correctly. Tip: put IPA in slashes right after the word and frame it so the clarification sounds intentional.

Further reads the model is built for: "[sarcastically] Oh, you again... [sigh]", "[door slams] Who's there?", or a calm "[measured] audiobook narration" across a full chapter.

How many languages does Eleven v4 support?

More than 90, and any voice can speak any of them.

How do I control emotion?

Add audio tags in square brackets, like [excited] or [whispers], inline in your text.

What is the difference between Eleven v4 and v4 Turbo?

Turbo keeps v4's expressive range at low latency for real-time agents; v4 gives the fullest performance for content.

Does it support SSML?

No. Use audio tags and punctuation instead of SSML.

Can a cloned voice speak other languages?

Yes, and by default it uses a native accent in the target language.

What is the maximum input length?

Up to 10,000 characters per generation.