Elevenlabs Launches Eleven Realtime Speech-to-Speech Models for Low-Latency Conversations

Voice AI startup Elevenlabs has introduced a new suite of speech-to-speech models called Eleven Realtime, engineered specifically to handle natural, bidirectional voice interactions with minimal latency. The release features two distinct versions: Eleven Realtime v1 Max and Eleven Realtime v1 Mini.

High Accuracy and Sub-Second Latency

The flagship variant, Eleven Realtime v1 Max, achieved a score of 98.8 percent on the Big Bench Audio benchmark alongside a response time of 0.57 seconds.

According to Elevenlabs, data from the Artificial Analysis Speech to Speech leaderboard places v1 Max near the top left corner of its chart, indicating a leading combination of high accuracy and low latency. The company stated that no other model currently tracked on the leaderboard is both faster and more accurate.

Specialized Roles: v1 Max vs. v1 Mini

Elevenlabs is positioning the two models for different enterprise and developer use cases:

• Eleven Realtime v1 Max: Tailored for complex interactive workflows, such as advanced customer support, where high precision and comprehensive reasoning are required.
• Eleven Realtime v1 Mini: Optimized for lower operational costs and fluid everyday conversation. The model includes conversational nuance detection, such as distinguishing whether a user has finished speaking or is simply offering brief conversational feedback like "mm-hmm."

Deployment and Pricing

Both models can be deployed through the ElevenAgents platform as well as via an OpenAI-compatible Realtime API, facilitating straightforward integration for developers already using standardized real-time protocols.

Pricing for both models is billed on a per-minute basis combined with token usage, with v1 Max priced higher than v1 Mini. Detailed tier breakdowns are published directly on the Elevenlabs pricing page.