Realtime Voice APIs for Game NPCs - 2026 Comparison

Hume has announced the shutdown of its EVI and TTS voice APIs, and every developer who built a talking character on them now has to move. It is a useful reminder that a realtime voice API is one layer of a game character, and that the layer can change under you.

Realtime voice models are good enough that a talking NPC demo takes an afternoon. Picking the API to build a game on is harder, because most of them were designed for customer service, not games. Here is how the main realtime voice APIs compare for game NPCs, and what you build on top of each.

Which realtime voice APIs can voice a game NPC?

OpenAI Realtime. Speech in, speech out over WebRTC or WebSocket, billed by audio and text tokens. A session ends at 60 minutes. There is no official Unity SDK, so the audio client and the token server are yours to build.

Gemini Live. Google's realtime API, also token-billed. A WebSocket connection lasts around 10 minutes, and session resumption lets a client reconnect and continue. Unity access goes through Firebase AI Logic, whose Unity SDK is in preview.

ElevenLabs Agents. A voice-first agent platform billed per minute, with the language model charged on top. Concurrency depends on your plan. ElevenLabs lists official SDKs for web and mobile frameworks, plus an early-stage Unity SDK that may change without notice.

Deepgram Voice Agent. Per-minute billing with options to bring your own language model or TTS. There is a C# SDK but no Unity package, and concurrency depends on your plan.

Amazon Nova Sonic. Speech-to-speech on Amazon Bedrock, token-billed, with connections limited to 8 minutes, after which you open a new connection and pass the history back in. It is aimed at contact centres and has no Unity SDK.

Hume EVI. Shutting down together with Hume's TTS API. If you shipped on it, plan the migration now, because voice IDs and configurations do not transfer to other providers.

How do they compare?

API	Status	Unity support	Pricing basis	Session limits
OpenAI Realtime	Active	No official SDK	Audio and text tokens	60-minute sessions
Gemini Live	Active	Firebase AI Logic, in preview	Audio and text tokens	Around 10-minute connections, resumable
ElevenLabs Agents	Active	Early-stage official SDK	Per minute plus LLM	Plan-based concurrency
Deepgram Voice Agent	Active	No package, C# SDK	Per minute	Plan-based concurrency
Amazon Nova Sonic	Active	No	Tokens	8-minute connections
Hume EVI	Shutting down	No	Not relevant now	Migrate now

What does a voice API leave you to build?

None of these APIs is a game character. With any of them, you still build:

An engine audio client. Capture, encoding, playback and interruption in Unity or Unreal, including WebGL, where Unity's built-in microphone class does not work.
A backend that issues short-lived tokens, so secret keys never ship inside your game build where players can extract them.
The character. Personality, knowledge, memory and guardrails, written and maintained as prompts.
Session handling. Reconnecting before time limits, carrying context across connections, and recovering when a player's network drops mid-sentence.
Game integration. Turning function calls into actions, and feeding scene state back so the character knows what the player is looking at.
Cost control for per-minute or per-token billing, including what happens when a player leaves the microphone open.
Maintenance as models, prices and APIs change, are renamed, or shut down.

For a small team, items 1, 2 and 4 alone are weeks of work before the first line of dialogue sounds right.

What does a voice NPC cost to run?

Two billing shapes dominate. Token billing (OpenAI, Gemini, Nova Sonic) charges for audio in and audio out, so long replies and long histories cost more as a conversation grows. Minute billing (ElevenLabs, Deepgram) charges for connection time, so silence while a player reads a menu still counts unless you close the session.

Either way, cost tracks how much players talk, not how many players you have. A game where one NPC is a constant companion behaves very differently from a game where players chat with a vendor for thirty seconds. Model it before launch, with the language model added on top where it is billed separately.

Neocortex uses flat monthly plans with a credit allowance: Free (200 monthly credits, personal use), Starter (EUR 9, personal and academic use), Pro (EUR 29, commercial use) and Team (EUR 99, five seats). Testing characters in the browser editor never uses credits.

How do you move off a voice API that is shutting down?

If you are leaving Hume EVI, or want a plan for the day another API changes, the move to Neocortex looks like this:

Export what you have. Download audio, transcripts and logs, and copy each character's system prompt and voice settings before access ends.
Rebuild each character in the editor. The system prompt splits into the Agent node (who they are), the Personality and Emotion nodes (how they react) and a Knowledge file (what they know).
Recreate the voice. Clone it from a sample you have the rights to, or describe it in words and use voice design.
Swap the client. Replace your WebSocket audio code with the Neocortex SDK for your engine, or the Web API for anything else.
Map tools to actions. Each function call you handled becomes an action your game executes when it appears in a reply.

Voice API or character platform?

Neocortex sits one level up. Instead of a voice API, you get the whole character: designed in a visual editor, voiced with a cloned or designed voice, and delivered through SDKs for Unity, Unreal and Godot (beta), or the Web API.

	Voice API	Neocortex
Engine SDK	Usually build your own	Unity, Unreal, Godot (beta)
WebGL microphone	Build your own	Included in the Unity SDK
Character design	Prompts	Visual node editor with Personality, Emotion, Knowledge and more
Voices	Provider presets	Voice cloning and voice design (5 on Pro, 20 on Team)
Game actions	Wire up function calls	Actions and interactables built in
Story structure	Build your own	Conversation Flows with `flowState` checkpoints
Vendor changes	Your problem	Our problem

The bottom line

Each of these APIs is one layer of a voice NPC, and the Hume shutdown shows how fast a layer can disappear. Neocortex is the whole character: streamed speech in and out, a voice of your own, actions your game executes, and engine SDKs that already handle the audio. Build one character on the free plan and hear it speak in your scene.

Individual comparisons

API details were last checked in October 2026.

Frequently Asked Questions

Is Hume EVI shutting down?

Yes. Hume is sunsetting its EVI and TTS APIs on November 13, 2026, and deletes account data after that date. Neocortex gives you streaming speech-to-text, text-to-speech and custom voices as part of a full character, with SDKs for Unity, Unreal and Godot (beta).

Does the OpenAI Realtime API have a Unity SDK?

No, OpenAI does not ship an official Unity SDK, so you build the audio client, token server and reconnect logic. Neocortex has a Unity SDK for Unity 6 with six sample scenes and microphone support in WebGL builds.

How long can a realtime voice session last?

It varies: OpenAI Realtime sessions end at 60 minutes, Gemini Live connections last around 10 minutes unless resumed, and Nova Sonic connections last 8 minutes. With Neocortex, the SDK handles the conversation and you send each player turn as it happens.

Which voice API is cheapest for game NPCs?

Token-billed APIs and per-minute APIs each scale with how much players talk, and the language model is often billed on top. Neocortex uses flat monthly plans with a credit allowance, and testing characters in the editor is free.

Can I use a custom voice for an NPC?

Most voice APIs offer preset voices, and cloning is often a separate product. On Neocortex you can clone a voice from a short sample you have the rights to, or design one from a text description: 5 custom voices on Pro and 20 on Team.