Cartesia brings sub-100ms voice AI to AWS users

Cartesia's approach addresses a fundamental tension in voice AI: traditional transformer-based text-to-speech models process entire sequences before producing output, so higher quality often means higher latency. State Space Models process sequences step by step with a constant memory footprint, enabling true streaming without architectural modifications. Sonic leverages this to begin generating audio before the full text input is received, and Ink uses native turn detection to identify when a speaker has finished, reducing the need for separate voice activity detection. Together, these models aim to deliver the responsiveness users expect from human conversation, and their availability on AWS—through a managed model package or self-managed containers—gives teams options for balancing operational control with time to deployment.