Skip to main content
Build → Voice Lab lets you hear how your AI Agent will sound. Enter a short script, choose a voice service and voice, play a sample, and compare two choices with A/B Testing. A voice service turns written words into spoken audio.

​
Preview a voice

1

Open Voice Lab

From a project, select Build → Voice Lab. Wait for the project voice configuration to load before choosing a provider.
2

Enter a representative script

Use a short sentence that includes the names, numbers, abbreviations, or language your agent will say in production. The page’s script field is labeled Enter a short script to preview this voice….
3

Configure the sample

Choose a TTS provider, model, language, and voice. If the provider supports it, set the model-specific voice options or enter a custom voice ID.
4

Generate and listen

Run the test and listen for pronunciation, latency, pacing, and natural turn endings. Repeat with a few scripts rather than choosing from one sentence alone.

​
Compare two configurations

Enable A/B Testing when you are deciding between providers, models, or voices. Configure Voice A and Voice B with the same script, generate each sample, and compare intelligibility, emotional fit, speed, and cost/latency trade-offs. A comparison is only useful when the text and playback conditions are the same.

​
Settings that matter

  • Provider and model determine the synthesis engine, available languages, voice catalog, and model-specific controls.
  • Voice selects a catalog voice. A Custom Voice ID can be used when your provider exposes a custom or cloned voice.
  • Language affects pronunciation and the language-specific catalog.
  • Speed, volume, and emotion are available only when the selected provider/model supports them; keep them consistent with the tone in the workflow prompt.
Voice Lab previews audio; it does not replace an end-to-end call test. The final behavior also depends on speech-to-text, the LLM, interruptions, welcome-message settings, and the workflow’s handover and timeout rules.

​
Move from preview to workflow

After choosing a configuration, open Workflows, select the relevant Voice Agent node, and update the Voice tab. Review the STT, TTS, and LLM settings together. Set Who Speaks First, the optional welcome message, interruption behavior, and node timeout in the General tab where applicable. Then run the full flow in Playground, connect it to a test inbound or Web Voice channel, and verify the call transcript in Operate → Conversations.

​
Troubleshoot preview failures

  • No providers or voices — wait for configuration to load and confirm the project has access to voice settings.
  • Provider loads but playback fails — check that both a provider and voice (or custom voice ID) are selected, then retry with a shorter script.
  • Voice sounds wrong in production — confirm the workflow uses the same provider, model, voice, language, and voice options that you tested in Voice Lab.
  • The new voice is not live — save and publish the workflow attached to the deployed voice surface.