Voice Architecture Patterns

The Voice AI Landscape

Where does your agent sit?

Three architectures can all produce a voice assistant. The difference is what travels through the middle, and whether your agent remains the brain.

Track the orange box. It is always your agent.

1

Cascade Sandwich

Three strangers. Your agent is the brain. You own every seam.

signal
STT: Deepgram Speech becomes text
text
YOUR AGENT The brain
text
TTS: Piper Text becomes speech
signal
2

Speech-Language Model

One unified model wraps your agent. The middle is still text.

Hume EVI: one model
signal
Hears + scores emotion Speech and expression
text + emotion
External, yours YOUR AGENT The brain
text
Re-voices Same model, other side
signal
3

Speech-to-Speech

One mind, in audio. No text in the middle.

signal
OpenAI Realtime audio never becomes text
signal
optional tool call
YOUR AGENT External brain, if needed