Models and architecture
Every Wirevox voice agent uses one of two pipeline architectures: Modular or Realtime. The architecture determines how speech is processed, which models and voices are compatible, and which conversation controls can run.
The best choice is not simply the newest model. Choose the smallest configuration that passes the agent's real call tests at an acceptable cost and latency.
In this topic
- Compare Modular and Realtime — understand the execution path, tradeoffs, and effect of switching architecture.
- Choose a model — review every currently selectable model and use a repeatable selection process.
- Review configuration compatibility — see which controls are native, handled by Wirevox, added as guidance, or inactive.
A practical starting point
Start with Modular + GPT-4o Mini, the current default for new modular agents, when you need broad voice choice, separate speech controls, or bilingual voice switching.
Start with Realtime when native speech-to-speech behavior is a core requirement and a compatible provider voice is acceptable.
These are starting configurations, not universal winners. Test both architectures for workflows where interruption quality, action reliability, or spoken latency materially affects the result.
What is stored in the draft
Wirevox keeps separate model and voice selections for Modular and Realtime. Switching the architecture changes which saved pair is active. The inactive pair is preserved so you can compare configurations without rebuilding it from scratch.
Architecture, model, voice, and related settings are draft configuration. Browser testing can use the draft, but live calls continue using the published version until you publish the change.
Before changing a live agent
- Record the current architecture, model, voice, estimated credits per minute, and test results.
- Change one major variable at a time.
- Resolve compatibility notices in the editor.
- Run the same representative test set.
- Compare transcripts, action results, interruptions, latency, and cost.
- Publish with a version name and note that identify the model change.
- Monitor live conversations before making another large change.