ElevenLabs Multilingual V2

Acerca de ElevenLabs

Fundado en 2022 · London, UK / New York, USA

ElevenLabs was founded in 2022 by Piotr Dabkowski and Mati Staniszewski, two Polish friends with backgrounds at Google and Palantir respectively. The product mission was to fix the poor dubbing experience for non-English films by producing AI voices with realistic intonation and emotion across many languages. The company is headquartered in London and New York with engineering centres in Warsaw and the Bay Area, and has raised over $280M across rounds led by Andreessen Horowitz and ICONIQ at valuations rising from $100M (Series A, 2023) to $3.3B (Series C, 2025). Multilingual V2, released in August 2023, became the default model for production audiobook and dubbing workflows at companies like Storytel, TheSoul Publishing and many indie publishers, and remained the flagship until v3 was previewed in 2025.

Arquitectura

Proprietary autoregressive Transformer TTS with neural codec

ElevenLabs Multilingual V2 is a hosted autoregressive Transformer text-to-speech model that predicts neural-codec audio tokens conditioned on text and a speaker embedding. The speaker embedding is obtained either from a stock voice (the curated 'Voice Library'), an Instant Voice Clone (1 minute of reference audio) or a Professional Voice Clone fine-tuned from 30 minutes of clean recording. Multilingual V2 supports 29 languages with strong code-switching between them in the same paragraph. Output is 24/44.1 kHz MP3 or PCM through a hosted API and the ElevenLabs Studio editor. The model is the production workhorse for ElevenLabs' Dubbing Studio (auto-translate plus voice match) and the AI Audiobook product. ElevenLabs has not published a technical paper, but the system architecturally resembles published neural-codec language-model TTS such as VALL-E and Voicebox.

Parámetros

Undisclosed

Contexto

5000 tokens

Capacidades

• 29 languages with seamless code-switching
• Voice Library with thousands of community and stock voices
• Instant Voice Clone (~1 min audio) and Professional Voice Clone (~30 min)
• Stability and similarity sliders for per-request prosody control
• Studio editor for multi-paragraph long-form projects
• Dubbing Studio with automatic translation and voice matching
• Up to ~5,000 characters per request
• Best for: production audiobooks, multilingual dubbing, content localisation

Entrenamiento y licencia

Not disclosed. Mix of licensed professional voice talent, public-domain audiobooks and opt-in user voice contributions.

Licencia: Proprietary commercial SaaS. Commercial use permitted on paid plans; customer voice clones remain customer property.

Pruebas de seguridad: Free AI Speech Classifier; mandatory consent statement and identity verification for Professional Voice Clones; watermarking; restricted access for likeness of named public figures.