Dear folks, I often have the problem of “very slow and robotic voices” here. After a certain point, the voice becomes completely distorted. there is no clear pattern when this happens. Sometimes 10 calls everyhing is fine and sometimes I have 5 calls in a row where this happens. Model used: GPT5 oder 5.1 Transscriber: Deepgram or AssemblyAI is used as the transcriber. The agents’ language is German. Voices: 11labs or Cartesia Logs: The logs show no errors at all. Token size is currently at 1250. Call id: 019aea11-1582-7000-82e3-07d38296a372 Starting at around 1m2s Any suggestions are welcome. Currently we need to take our VAPI Agents offline as this is unacceptable. Could it be the latency of the llm oder something VAPI internally. I can provide recordings for sure.

Suggestion: Try switching your voice provider to Azure or OpenAI TTS for your German agents, as this often resolves "robotic" or distorted speech issues. In your Vapi dashboard, go to your assistant's Voice section and select a different provider (e.g., Azure), then save and test:

# Example: Change voice provider in dashboard
1. Navigate to Assistants > [Your Assistant]
2. Under Voice, select Provider: Azure
3. Choose a German voice (e.g., de-DE-KatjaNeural)
4. Click Save and test a call

This quick change can help you determine if the issue is provider-related. For more details, see the [debugging documentation](https://docs.vapi.ai/debugging). Source: - [Debugging Documentation](https://docs.vapi.ai/debugging)

Hi wh1stler, For your issue with slow and robotic voices with your German agents, consider these steps: 1. **Change Voice Provider**: - Try switching to Azure TTS or OpenAI TTS to reduce robotic or distorted speech, as they often provide more stable performance. You can follow these steps:

markdown
     1. Navigate to `Assistants > [Your Assistant]` in your Vapi dashboard.
     2. Under Voice, select Provider: Azure.
     3. Choose a German voice (e.g., `de-DE-KatjaNeural`).
     4. Click Save and test a call.

2. **Verify Provider Status**: - Check the status pages of ElevenLabs or Cartesia to ensure there are no ongoing service issues that could affect voice output. 3. **Optimize Model Settings**: - Consider reducing the model's temperature and adjusting

max tokens

to maintain output consistency while also keeping token size optimized. 4. **Minimize Latency**: - If GPT-5 or 5.1 is causing noticeable delays, try using a different model with similar capabilities but potentially better latency performance. 5. **Review TTS Settings**: - Confirm the TTS setup meets the PCM requirements: mono, 16-bit little-endian, with matching sample rates, as incorrect formats often lead to playback issues. 6. **Fallback Providers**: - Configure a fallback voice provider to switch automatically if your primary provider fails. By following these steps, you should be able to diagnose whether the issue is provider-related or something that requires further investigation on the configuration side. If challenges persist, reviewing real-time transcripts via the Vapi dashboard might also provide additional insight. For more help, consult the [debugging documentation](https://docs.vapi.ai/debugging).