Building ultra-low-latency voice AI applications on AWS

Amazon's post explains how Cartesia, an AWS Partner, uses State Space Models and AWS infrastructure to run real-time voice AI with sub-100ms text-to-speech and speech-to-text.

• Cartesia's two production models are Sonic for speech generation and Ink for transcription, both built on State Space Model architecture rather than the transformer attention mechanism.
• Transformers process entire sequences before producing output, so higher quality often means higher latency, while SSMs carry a compact internal state forward and process input token by token.
• Sonic returns a first audio byte in approximately 90ms, supports over 44 languages, and can clone a voice from 10-second audio samples.
• Ink adds native turn detection, context-aware end pointing, streaming partial transcripts, noise robustness and multi-speaker support, so a downstream LLM can start forming a response before the speaker's turn ends.