Voices, languages, and speech
A Wirevox voice agent has three related speech decisions: how caller audio becomes text, which language the conversation uses, and how the response becomes audio. The available controls depend first on whether the agent uses the Modular or Realtime architecture.
- Modular separates speech recognition, the language model, and text-to-speech. You can choose the speech recognition provider independently, use voices from several providers, and configure a secondary language and voice.
- Realtime sends the conversation through one provider-owned speech-to-speech session. The provider controls recognition and synthesis, and the session uses one compatible voice.
Changing a voice does not change the system prompt or model. Changing the architecture can change which speech controls are executed.
In this topic
- Choose a voice — compare provider compatibility, preview voices, set speed, and create a private custom voice.
- Configure languages and bilingual behavior — choose the primary language and understand Modular language switching.
- Configure speech recognition — choose a Modular transcription provider, patience level, and recognition terms.
- Configure voice delivery — control personality, expressiveness, reactions, pitch, pronunciation, and provider-native options.
- Test and troubleshoot speech — evaluate real calls and diagnose recognition, language, voice, and delivery problems.
Configuration boundary
| Capability | Modular | Realtime |
|---|---|---|
| Independent speech recognition provider | Yes | No; owned by the Realtime provider |
| Voices from a separate TTS provider | Yes | No; voice must match the Realtime provider |
| Primary language | Yes | Yes |
| Configured secondary language and voice | Yes | Not currently executed |
| Recognition key terms | Passed to supported STT providers | Native or guided, depending on provider |
| Turn-taking profile | Derived through Modular recognition controls | Provider-native |
| Vocal-mood awareness | Not reliable; the model receives transcript text | Available to the Realtime model |
The editor identifies controls as native, hybrid, pipeline-managed, guided, or inactive. Treat those labels as execution information, not decoration.
Current selectable languages
The voice-agent editor currently exposes English and French because those languages have configured voice options in the live catalog. Do not promise a language based only on a model provider's general language claims. It must be selectable and testable in the Wirevox editor.
Draft and published behavior
Voice, language, recognition, and delivery changes update the agent draft. A browser test can use that draft, but assigned phone numbers continue using the published version until you publish the change.
Recommended sequence
- Choose the architecture and model.
- Set the primary language.
- Select and preview a compatible voice.
- For Modular agents, select the speech recognition provider.
- Add a secondary language only when the workflow requires it.
- Tune delivery after the prompt and action flow are stable.
- Test with representative speakers, devices, noise, names, dates, and interruptions.
- Review the estimated credit rate, publish, and monitor live conversations.