AWS walks through a full Pipecat voice agent with Deepgram on SageMaker

An AWS post co-written with Pipecat and Deepgram deploys a complete voice agent: Deepgram STT/TTS on SageMaker, Claude on Bedrock, Pipecat on ECS.

AWS published a step-by-step reference deployment (1 October 2026) for a production-style voice agent. It covers VAD, speech-to-text, LLM, text-to-speech and telephony in one streaming Pipecat pipeline. It is worth reading for the pipeline design alone, even if you never deploy on SageMaker. Pipecat's Kwindla Hultman Kramer, a co-author, flagged it on X and says Pipecat helped build the original SageMaker bidirectional reference implementations.

Per the post, the setup works like this:

The post says auto scaling keys on active calls rather than CPU, and a CloudWatch dashboard tracks latency and interruptions. Repository skills let Claude Code drive the deployment. Pipecat's docs confirm a DeepgramSageMakerSTTService for this path.

The costs are real. The post warns that the SageMaker GPU endpoints are the most expensive component and are billed per hour regardless of usage. It also says the ECS service is always on. It gives no end-to-end latency figure and no per-call cost, and whether anything scales to zero is unclear. For STT latency numbers you can check yourself, Pipecat maintains an open benchmark.

There are lighter-weight options:

Sources