Skip to main content
Voice controls the speech callers hear during phone conversations. The selected voice and its settings belong to the agent and apply across its inbound and outbound flows. Use Guidelines to control the agent’s personality, tone, and response style. Voice controls how those responses sound when spoken.

​
Choose a voice

Open Agent Design, then select Voice.
  1. Play a voice to preview it with the sample text in the test card.
  2. Search by voice name or filter the library by language, gender, accent, or other available traits.
  3. Select the voice you want the agent to use.
Voice library showing the selected voice, filters, preview card, and Voice Settings panel
The selected voice appears first in the library. Use the heart control to save favorites, and use Mine to view voices owned by you or your organization. Choose a voice that remains clear at phone-call audio quality and fits the conversations the agent handles. Accent, pacing, and pronunciation are usually more important than how expressive a voice sounds in a short sample.

​
Customize the voice

Use Voice Settings to adjust the selected voice. Changes save automatically. Changing to a voice without multi-lingual support turns Multi-Lingual off. Re-enable it after selecting a voice that supports it.

​
Add pronunciations

Pronunciations is available for Cartesia voices. Add an entry when the agent needs to say a name, brand, acronym, or specialized term consistently. Enter the written word and its pronunciation, then play the entry before saving it. Expand Phonemes only when the generated pronunciation needs a precise adjustment.

​
Configure transcription

Expand Advanced only when the default speech recognition needs adjustment:
  • Transcriber selects the service that converts caller speech into text.
  • Transcriber Languages prepares supported transcribers to recognize selected languages. This setting is available for Gladia and Soniox.
  • Transcription Keywords help the transcriber recognize uncommon names, product terms, or identifiers.
Keep the default unless testing shows that another transcriber better fits the language or audio conditions of your calls.

​
Test the voice

Enter a representative sentence in the test card. Include names, numbers, or specialized terms the agent will commonly say.
  1. Preview several voices with the same text so the comparison is meaningful.
  2. Select a voice, adjust its settings, and preview it again.
  3. Confirm the pronunciation, pace, and background sound before testing the complete agent.
  4. Use Web Call to hear the voice in a real conversation and check turn-taking, interruptions, and phone-call audio quality.
A voice preview tests synthesized speech, not the complete call experience. Always complete a Web Call before relying on a new voice or major setting change.

​
Add or clone a voice

Select Add or Clone a Voice when the built-in library does not contain the voice you need.
Choose Add a voice, then select the provider and model. Enter the voice ID from the provider and select its language. Adding the voice also selects it for the current agent, so test it before using the agent for live calls.
Choose Clone a voice, then upload or record a clear 1–2 minute sample. Enter a unique name, gender, and language, and confirm that you have permission to clone and use the recording. Compare the generated samples before saving one. The saved voice becomes the current agent’s selected voice.

​
Troubleshooting

The selected voice does not support multi-lingual responses. Select a voice marked as multi-lingual, then enable the setting again in Voice Settings.
If the selected voice supports Pronunciations, add the word and test a clearer pronunciation. If the setting is unavailable, select a Cartesia voice or another voice that already pronounces the term correctly.
Confirm that the intended voice is still selected and rerun the preview after changing Speed, Background Noise, or Pronunciations. Then use Web Call: phone audio, caller interruptions, and the surrounding conversation are not represented by an isolated voice preview.
Last modified on July 17, 2026