ElevenLabs unveils voice AI supporting 90 languages

[Financial News] ElevenLabs has unveiled an AI voice agent model that integrates voice conversations and reasoning capabilities. The move aims to simplify the existing approach, which required combining separate technologies for everything from speech recognition and conversational flow control to speech synthesis, and ease the burden on businesses and developers of building voice AI services.
According to industry sources on the 3rd, global AI audio company ElevenLabs has launched Eleven Realtime, a speech-to-speech (STS) model. It comes in two versions: Eleven Realtime v1 Mini, focused on fast, natural conversations, and Eleven Realtime v1 Max, specialized in handling complex tasks. The models are available through Eleven Agent, a voice chat agent platform, and a real-time application programming interface (API).
Previously, building a natural-sounding voice agent required selecting and combining multiple technologies separately, including speech recognition, voice activity detection, conversational turn-taking, large language models (LLMs), and speech synthesis. Eleven Realtime integrates these capabilities into a single system, allowing users to build a voice agent simply by selecting a model.
Both models support more than 90 languages, including Korean and Japanese, and users can choose a voice from a library of more than 17,000 voices owned by ElevenLabs. They also support natural conversational turn-taking, responses to user interruptions, and external tool calls.
The Mini model is suited to services requiring quick responses, such as AI assistants, educational tutors, and characters for games and entertainment. The Max model focuses on handling multistep tasks, including customer support, lead qualification, and retrieving information using external tools. Users can switch between the two models simply by changing their selection to suit the purpose of their service.
In an evaluation by third-party evaluator Artificial Analysis, as of October this year, the Max model scored 98.8% on Big Bench Audio, which measures voice-based reasoning ability. It took 0.57 seconds to begin answering after a question ended. On Full Duplex Bench, which measures responses to silence, turn-taking, and interruptions during conversation, the Mini model scored 95.8% and the Max model 94.1%.
ElevenLabs says it also plans to support the use of voice AI by Korean businesses and developers. As the need to make voice-based operations, such as customer service, more efficient grows with the transition to a super-aged society, the company is expected to target the enterprise voice agent market by emphasizing multilingual support, including Korean, and simpler model development.
A representative of ElevenLabs said, “Through Eleven Realtime, which supports more than 90 languages, including Korean, we plan to help Korean businesses and developers build voice agents that can converse naturally with customers.”
[email protected] Jang Min-kwon Reporter