Just wondering what the thinking is at VAPI and within the dev community over how to integrate and work with the future of voice ai conversational models like Gemini 3.1 Flash where audio to audio and is going to be standard? For example we're currently sending a 'post next best action' to our AWS server to keep an agent on track through a more complex qualification call in the legal space. So being able to keep with that but do away with the STT + LLM + TTS stack with a single model that handles everything and hook it up with VAPI's config would be game changing. Thoughts? Timeline to market?

VAPI is already moving toward integrated audio-to-audio conversational AI by supporting Gemini 2.0 models with real-time configuration and customizable speech output. To experiment with this, you can set up your assistant using the following config snippet:

json
"model": "gemini-2.0-flash-realtime-exp",
"realtimeConfig": {
  "speechConfig": { "voice": "Puck" },
  "voiceConfig": { /* additional options */ }
}

While Gemini 3.1 Flash isn't mentioned yet, this approach positions you to adopt future audio-to-audio models as soon as they're available in VAPI's platform. Reference: See the [VAPI changelog for Gemini 2.0 integration details](https://docs.vapi.ai/changelog/2025/1/7). Source: - [Changelog](https://docs.vapi.ai/changelog/2025/1/7)

You’re right audio-native will cut latency, but keeping control like your NBA logic is the tricky part. Most teams still keep orchestration in Vapi or Amazon Web Services and let audio models handle realtime flow. Hybrid setups feel like the safe move for now, especially for legal calls. Is your NBA driven more by transcripts or by state/intent signals? @Magneto

Message marked as helpful by @Magneto! 🎉

Hey Matt... more by state and intent signals.