Cartesia's SSM voice AI hits ~90ms latency, deployable in your own AWS VPC
AWS ML Blog · rss · 2026-11-04
An AWS blog post covers how partner Cartesia uses State Space Models (SSMs) to deliver sub-100ms real-time voice AI on AWS. Traditional transformer TTS must process full sequences before output, forcing a quality-vs-latency tradeoff; SSMs process sequences step by step with a compact fixed-size state, avoid quadratic compute growth, and natively support streaming.
Two models:
- Sonic (TTS): 90ms first-byte latency, 44+ languages, voice cloning from 10-second samples, and audio generation begins before full text input arrives.
- Ink (STT): native turn detection that uses semantic understanding to distinguish mid-sentence pauses from true turn boundaries (no separate VAD needed), streaming partial transcripts so downstream LLMs can start early, multi-speaker support, and noise robustness.
Deployment: Sonic ships as a SageMaker AI model package via AWS Marketplace, deployable to an in-VPC endpoint (ml.g6e.xlarge, ml.g7e.2xlarge, ml.p5.48xlarge, etc.) with inference data staying in your account; Ink is currently available via the Cartesia API billed through Marketplace. Target use cases include contact centers, telehealth, gaming NPCs, accessibility, and AI sales agents.
More from Infra
- SurrealDB 4 launches with SurrealDB Earth: serverless agent database on Cloudflare — ritakozlov · 2026-11-04
- The hidden latency inside your database: PlanetScale on network-attached storage costs — DanielLockyer · 2026-11-04
- Binance discloses $6B+ heist, largest in history, warns of AI-accelerated social engineering — evilsocket · 2026-11-04
- US now spends twice as much building data centers as offices: $96.5B vs $46B — ohlennart · 2026-11-04
- Jensen Huang and DHH discuss inference and open-weight models at Fireworks Forge 2026 — gethackteam · 2026-11-04
- AIOps platform cuts agent token costs 30%+ with Pydantic's open-source sandbox Monty — samuelcolvin · 2026-11-04