Cartesia Releases Managed Agents, Ranking First in Blind Human Preference Tests on Speech Agent Arena
Cartesia has launched Managed Agents, an end-to-end voice agent platform for enterprises, which ranks first in blind human preference tests on Artificial Analysis's Speech Agent Arena and first on ServiceNow EVA-Bench. The platform uses a cascaded architecture, jointly tuning its proprietary voice models and the agent harness as a single system. Its H-Nets learn directly from raw bytes, achieving 3–5x efficiency gains in controlled scaling experiments, and at larger scale matches an open-source model trained on 11 trillion tokens with roughly 10x less training data and compute.
Why read this
AI score 54
Cartesia's cascaded voice agent outperforms native speech-to-speech models in blind human preference tests, giving readers insight into the architectural choices and evaluation results of voice agents.
Topics
View original postIntroducing Cartesia Managed Agents: our end-to-end voice agent platform for enterprise. #1 on @ArtificialAnlys Speech Agent Arena and @ServiceNow EVA-Bench, built by co-designing our voice models and harness as one system. Check it out → https://t.co/7kJhCYqjQo https://t.co/NVOAds9czf
View original
Introducing Cartesia Managed Agents: our end-to-end voice agent platform for enterprise. #1 on @ArtificialAnlys Speech Agent Arena and @ServiceNow EVA-Bench, built by co-designing our voice models and harness as one system. Check it out → https://t.co/7kJhCYqjQo https://t.co/NVOAds9czf
Body
A cascaded voice agent just won a blind human-preference ranking against native speech-to-speech models.
Today, Cartesia launched Managed Agents, ranked #1 on Artificial Analysis’ Speech Agent Arena.
Most LLMs learn from chunks of text picked by a tokenizer. But here, H-Nets learn directly from raw bytes and figure out those boundaries themselves.
This leads to 3–5x better efficiency in controlled scaling experiments. At larger scale, matching an open-source model trained on 11 trillion tokens with roughly 10x less training data and compute.
> #1 in blind human preference and #1 in enterprise accuracy > Cascaded stack that sounds as natural as speech-to-speech > Own speech models and agent harness tuned as one system