ElevenLabs launches two real-time speech-to-speech models

By Splitfeed EditorialLast updated Oct 20, 2026, 19:10 UTC

On October 20, 2025, ElevenLabs released two new speech-to-speech models under the name Eleven Realtime, built specifically for real-time AI conversations. The more capable of the two, the v1 Max variant, recorded a score of 98.8 percent on the Big Bench Audio benchmark and achieved a response latency of 0.57 seconds. ElevenLabs stated that, as of the release date, no model listed on the Artificial Analysis leaderboard combined a faster response time with higher accuracy than its new offering.

The release matters because latency and comprehension accuracy are the two variables that most directly determine whether a voice AI system feels natural in a live conversation. A sub-600-millisecond response time sits close to the range of ordinary human conversational pauses, and a benchmark score near 99 percent suggests strong audio understanding. The genuine dispute here is whether the Big Bench Audio benchmark and the Artificial Analysis leaderboard together constitute a complete picture of real-world performance, or whether they capture only the conditions ElevenLabs chose to highlight. Third-party testing under varied acoustic environments, accents and use cases would be needed to confirm the broader claim.

None of the three camps, Pro-AI, Anti-AI or Middle Ground, had published reactions at the time of writing. Pro-AI voices would typically celebrate a release like this as evidence that voice AI is reaching conversational quality, pointing to the benchmark figures as validation. Anti-AI voices would likely question what the performance gains enable in practice, raising concerns about synthetic voice misuse and whether speed improvements make deceptive audio harder to detect. Middle Ground observers would probably call for independent evaluation of the benchmark claims before drawing conclusions about market leadership.

The figure to watch is how the Eleven Realtime models perform when independent researchers or developers run evaluations outside ElevenLabs' own framing, particularly on the Artificial Analysis leaderboard as competing labs update or release their own models in response.