About other protocols and frameworks.

Hi. I was wondering if other protocols and frameworks can be done similarly. My use case requires real-time inference on audio chunks. I tried using real-time endpoints(Multi-model, PyTorch) and the latency(~0.8S) doesn't allow real-time inference(~6ms-25ms) on audio chunks from a microphone. I am looking for some references that demonstrate the usage of socket servers in SageMaker. If SageMaker doesn't allow that, is there a preferred way for deploying such ML workflows in AWS? TIA