Artificial Intelligence
ElevenLabs launches Eleven v4 speech model
ElevenLabs released Eleven v4 and a Turbo variant, saying they improved control of speech cues, voice consistency and language coverage while Turbo was built for low-latency voice agents.
- ElevenLabs released Eleven v4 and Eleven v4 Turbo for speech generation.
- The company said v4 handled more than 90 languages, accepted up to 10,000 characters per request and improved voice cloning.
- ElevenLabs said Turbo began producing speech in about 150 milliseconds and was designed for real-time voice-agent use.
- The article said ElevenLabs set standard API prices at $80 per million characters for v4 and $40 for Turbo, with lower introductory rates until 12 October.
- The article said v4 ranked ahead of Cartesia Sonic 3.6 and Google’s Gemini 3.8 Flash TTS on Artificial Analysis’ Voice Arena leaderboard.
ElevenLabs released Eleven v4 and Eleven v4 Turbo for speech synthesis. The company said the main model followed direction cues more closely and kept voices stable across longer productions.
ElevenLabs said users could guide output with tags or plain sentences, and use phonetic spellings for names and technical terms. It said the model handled up to 10,000 characters at a time and supported more than 90 languages, up from about 70 in v3.
The company said Turbo was intended for live uses such as customer service and game characters, and that it began audible speech in about 150 milliseconds. In ElevenLabs’ tests, that was faster than Cartesia Sonic 3.6 and OpenAI’s GPT-4o mini TTS.
The article said ElevenLabs placed v4 ahead of Cartesia Sonic 3.6 and Google’s Gemini 3.8 Flash TTS on Artificial Analysis’ Provider Voice Arena leaderboard. It also said ElevenLabs’ blind tests found about three-quarters of listeners preferred v4 over models from Cartesia, Inworld and Google.
ElevenLabs said v4 was available in ElevenAgents, ElevenCreative and through its API. It listed standard API prices at $80 per million characters for v4 and $40 for Turbo, with temporary reductions until 12 October, and said customer data was stored in the US by default.
Named in this story
People
- Michael Caine
- was cited as an example of a celebrity voice in the marketplace
Companies
- ElevenLabs
- released the speech models and supplied the performance claims
- Cartesia
- made Sonic 3.6, a model used in comparisons
- OpenAI
- made GPT-4o mini TTS, a latency comparator
- made Gemini 3.8 Flash TTS, which appeared in comparisons
- Inworld
- supplied a model used in ElevenLabs’ blind tests
Organisations
- Artificial Analysis
- ran the Voice Arena leaderboard cited in the article
Products and systems
- Eleven v4
- the main speech model that ElevenLabs launched
- Eleven v4 Turbo
- the lower-latency variant aimed at live voice use
- Eleven v3
- the earlier model used as the comparison point
- ElevenAgents
- the platform optimised alongside Turbo
- ElevenCreative
- one of the products where v4 was made available
- Cartesia Sonic 3.6
- a competitor model used in latency and ranking comparisons
- GPT-4o mini TTS
- a comparator in ElevenLabs’ timing tests
- Gemini 3.8 Flash TTS
- a comparator in leaderboard and preference tests
- Provider Voice Arena leaderboard
- the ranking system where v4 was placed ahead of rivals
Places
- US
- the default storage location for customer data
- EU
- one region where enterprise customers could isolate data
- India
- one region where enterprise customers could isolate data
- Singapore
- one region where enterprise customers could isolate data
How the source tells it
The piece read like a product launch write-up, leaning on the company’s own claims, benchmark comparisons and listener tests to present the release favourably.
- vendor boosterism the company’s claims were repeated throughout and left to stand as the main evidence
- competitive framing side-by-side latency figures, leaderboard placement and preference percentages were used to imply superiority