Building Game NPCs on the Gemini Live API vs Using Neocortex
Google's Gemini Live API is one of the most cost-effective ways to hold a real-time voice conversation with an AI model. For game developers, it is tempting: point a microphone at it, give it a persona, and you have a talking NPC.
Here is what that looks like in practice in 2026.
What the Gemini Live API gives you
From Google's documentation (checked October 2026):
- Audio-to-audio models over a stateful WebSocket connection, with a free tier and low per-token prices.
- Ephemeral tokens recommended for client apps, so you need a backend to issue them.
- Session limits. Audio-only sessions run 15 minutes unless you enable context compression, and connections last around 10 minutes, so you handle resumption.
- Unity through Firebase AI Logic, which is in preview, has no SLA, and supports only older Live models rather than the newest stable ones.
What you build yourself
- A Unity audio client, or adoption of the Firebase preview with its model limits.
- A backend for ephemeral tokens.
- Reconnection and session resumption.
- Character design as prompts, plus memory across sessions.
- Game integration for function calls, and guardrails.
- Cost monitoring for token-based billing.
Gemini Live vs Neocortex
| DIY on Gemini Live | Neocortex | |
|---|---|---|
| Unity | Firebase AI Logic preview, older models only | Unity SDK with voice samples |
| Other engines | Build your own client | Unreal, Godot (beta), Web API |
| Character design | Prompts in code | Visual node editor |
| Voices | Google's voices | Voice cloning and voice design |
| Sessions | Manage limits and resumption | Session handling built in |
| Game actions | Wire up function calls | Actions and interactable objects built in |
| Local AI | No | Local LLMs in Unity via Ollama |
Why teams choose Neocortex
Ship on stable tooling. The Neocortex Unity SDK is our core product, not a preview.
Your characters, your voices. Clone or design a unique voice for every character instead of choosing from a shared set.
Designers in control. Personality, emotion, knowledge, intents and actions live in a visual graph, not a prompt string.
No session plumbing. Memory, sessions and streaming are handled for you.
When Gemini Live is the right call
If you are building your own voice stack and want low raw model costs, Gemini Live is a strong engine to build on. If you want voiced NPCs in your game this week, Neocortex gets you there without the plumbing.
Google details were last checked in October 2026.
Written by
Sercan Altundas
Date
Oct 17, 2026
Read time
4 min
