ElevenLabs v3 (alpha)

Über ElevenLabs

Gegründet 2022 · London, UK / New York, USA

ElevenLabs was founded in 2022 by Piotr Dabkowski (CTO, ex-Google ML engineer) and Mati Staniszewski (CEO, ex-Palantir), two Polish high-school friends frustrated with the poor quality of TV-show dubbing in Polish. The company set out to build voice AI that captures intonation and emotion across languages. Headquartered in London and New York with engineering hubs in Warsaw and the Bay Area, ElevenLabs raised a $19M Series A in June 2023 led by Andreessen Horowitz, a $80M Series B in January 2024 also led by a16z at a $1.1B valuation, and a $180M Series C in January 2025 at a $3.3B valuation co-led by a16z and ICONIQ. ElevenLabs v3 (alpha) was previewed in 2025 as the next generation flagship model with expressive emotion tags, longer context and more languages, succeeding the Multilingual V2 family that became the de-facto standard for AI dubbing.

Architektur

Proprietary autoregressive Transformer TTS with neural codec and emotion/prosody conditioning

ElevenLabs v3 (alpha) is the company's 2025 flagship text-to-speech model and the first ElevenLabs system to expose explicit emotion and event tags inside text input ([whispers], [laughs], [angry], [sighs]). It is a proprietary Transformer-based autoregressive model that predicts neural-codec audio tokens conditioned on a text prompt and a speaker embedding obtained from a few seconds of reference audio (Instant Voice Clone) or a fully fine-tuned voice (Professional Voice Clone, requires ~30 minutes of clean audio). v3 expands language coverage from 29 (v2) to 70+ languages, lengthens the input window to roughly 10,000 characters per request, and adds dialogue mode for multi-speaker scenes. ElevenLabs has not published a technical paper; product blog posts describe internal improvements in speaker disentanglement, code-switching and emotional range. v3 is offered through the same hosted API and Studio UI as Multilingual V2 but at higher latency and price.

Parameter

Undisclosed

Kontext

10.000 Token

Fähigkeiten

• Expressive emotion and event tags ([laughs], [whispers], [angry], [crying])
• 70+ languages with high-quality code-switching
• Multi-speaker dialogue mode for podcast and audiobook generation
• Instant Voice Clone from ~1 minute of audio and Professional Voice Clone from ~30 minutes
• Long-form input up to ~10,000 characters per request
• Studio editor for multi-paragraph projects with per-line speaker control
• Best for: audiobooks, dubbing, narrative podcasts, character voices for games

Training & Lizenz

Not disclosed. ElevenLabs licences professional voice talent, uses public-domain audiobooks and crowd-sourced opt-in voice contributions; commercial recordings are excluded per their public statements.

Lizenz: Proprietary commercial SaaS. Commercial use of generated audio is permitted on paid plans; voice clones remain customer property.

Sicherheitstests: AI Speech Classifier offered for free; mandatory consent statement and KYC for Professional Voice Clones; inaudible watermark on outputs; mis-use bans after 2024 deepfake incidents.