Self-hosting an AI voice stack: a practical guide

Running voice AI on your own infrastructure gives you control over data, latency, and cost. It used to mean stitching together a SIP server, a media server, a speech stack, and an agent framework. With DialPress it is one stack with one set of deploy options.

Start with one server

A single Linux installer sets up the gateway, speech, and agents as hardened systemd services with TLS for your domain. It is the fastest way to a working phone number that an AI agent answers.

Move to containers

Docker Compose runs every module as its own container, with memory and CPU bounds per service. Scale a busy module on its own instead of scaling everything.

Grow into Kubernetes

The Helm chart runs each module as its own workload, autoscaling on live metrics. Stopping pods drain first, so deploys and scale-downs never drop a call.

Put speech on GPUs

Speech recognition and agents can run on GPU machines over mutual TLS, balanced by model and load. Weight bigger GPUs higher; route each model only to the machines that serve it.

Keep the same code

Your routes, agents, and apps do not change between a laptop and a fleet. That is the point of an infrastructure layer.