Incorrect voice output
Wrong accent or language, mispronounced words, a missing or rejected voice, and an agent that speaks the wrong thing.
What you will achieve
You will correct how the agent sounds and what it says aloud.
Symptoms
| Symptom | Likely cause | Fix |
|---|---|---|
| The accent is wrong, or English with a foreign accent | An OpenAI voice with a non-English language (OpenAI voices are tuned for English) | Choose a Cartesia or ElevenLabs voice for that language |
| The voice list changed after switching provider | Each provider has its own voices | Choose a voice again after switching |
| A pasted voice ID does not work | Wrong provider's ID, or an ElevenLabs library voice not in My Voices | Add the voice to My Voices in ElevenLabs first |
| Urdu sounds poor or fails | Urdu needs specific models | Use ElevenLabs (which uses its v4 Turbo model for Urdu) or Cartesia; compare by listening |
| Mispronounced names | Unusual spelling | Spell them phonetically in the Instructions |
| The agent reads out symbols or a web address | The model wrote text meant for reading | Tell it in Instructions to speak naturally and spell addresses aloud |
| The agent says things it should not | The prompt is too loose | Tighten the Instructions; lower Temperature |
| The language changes mid-call | The model followed the caller's language | Add "Always reply in <language>" to Instructions |
Check by listening. Audio quality for each language, Urdu most of all, has not been judged by listening. Always run Talk with a native speaker before publishing.
Testing
Use Talk, ask for a price, a date and a phone number, and listen to how each is spoken.
Expected result
Numbers, dates and names sound natural in the chosen voice and language.
Related guides
Something here does not match what you see in the dashboard? Let us know.