Hi, no matter how i try the assistant does never speaks first when i set it via "First Message Mode". It either waits for the user to speak then speaks the first message or when on the auto-generated message starts to behave in a non predictable way, can speak in any language, or just halluciante a conversation. Someone one elses's issue: #1328417131011772457

To ensure your assistant speaks first, set the following properties in your assistant configuration:

json
{
  "firstMessageMode": "assistant-speaks-first",
  "firstMessage": "Hello! How can I help you today?"
}

Make sure

firstMessage

is set; otherwise, the assistant will wait for the user to speak. If issues persist, double-check that these fields are correctly passed in your API request and that no conflicting settings are present. Reference: See the "firstMessageMode" and "firstMessage" options in the [Vapi documentation](https://docs.vapi.ai/api-reference/assistants/create?explorer=true). Source: - [Vapi API Reference: Create Assistant](https://docs.vapi.ai/api-reference/assistants/create?explorer=true)

Ok, please help me understand how to work with vapi, since the CLI is not working for me: #1423000508498579508 Should i create custom wrapper for the VAPI api? what do you suggest?

Hi Liam T Bilich, To address your inquiry effectively, it would be helpful to know more specifics about the setup or the issue you are facing. If your query is about configuring or optimizing voice interactions using Vapi's models or voice services, here’s a brief overview of critical settings you might consider: 1. **Model Selection**: Use the

gpt-realtime-2025-08-28

for production needs, supporting structured outputs and complex orchestration tasks. If cost or latency is a concern, you might opt for one of the preview versions like

gpt-4o-mini-realtime-preview-2024-12-17

[OpenAI Realtime Documentation](https://docs.vapi.ai/openai-realtime). 2. **Voice Configuration**: For real-time models, ensure you are selecting compatible voices like

marin

or

cedar

, which are optimized for these setups. Configurations can be defined in JSON format:

json
   {
     "voice": {
       "provider": "openai",
       "voiceId": "marin"
     }
   }

3. **Custom Vocabulary and Thresholds**: Enhance speech recognition accuracy using custom vocabulary features from AssemblyAI or Speechmatics. Adjust end-of-turn thresholds for natural conversation flow with properties like

eotThreshold

and

eotTimeoutMs

[Changelog for Transcription Enhancements](https://docs.vapi.ai/changelog/2025/9/8). If you have specific symptoms or configurations you need help with, feel free to share more details.