What changed
AWS announced that Cartesia, an AWS Partner, is using State Space Models (SSMs) to deliver ultra-low-latency voice AI applications. These models provide sub-100ms text-to-speech (TTS) and speech-to-text (STT) capabilities, enabling real-time voice interactions. The solution is designed for use cases such as virtual assistants, contact centers, telehealth, gaming, and accessibility tools.
What this means in practice
Cartesia's models, Sonic for TTS and Ink for STT, are optimized for sub-100ms first-byte latency while maintaining studio-quality output. Sonic generates audio almost instantaneously, with a first audio byte in approximately 90ms. Ink supports native turn detection and streaming transcription, allowing downstream models to begin forming responses before a conversation ends. These models are available through AWS Marketplace, with Sonic deployed as an Amazon SageMaker AI model package and Ink accessible via the Cartesia API.
Limitations and availability
The supplied announcement does not specify exact performance benchmarks or independent testing results. Deployment of Sonic requires an Amazon SageMaker AI endpoint running within a virtual private cloud (VPC), with supported instance types including ml.g6e.xlarge or larger. Ink is currently available through the Cartesia API. The models are intended for developers seeking real-time voice AI solutions with low latency and high-quality output.
