Amazon Nova Sonic for Game NPCs vs Neocortex

Amazon has released Nova 2.5 Sonic, the newest of its speech-to-speech models on Bedrock, and it is now generally available. AWS says it improves reasoning, instruction following and tool-calling accuracy over Nova 2 Sonic, with lower latency, at the same price.

Nova Sonic is built for real-time voice agents, and the launch examples are customer service: an agent that looks up an order, checks return eligibility and starts a return in one call. So the question for game developers is a fair one. Can it power NPCs?

Who Nova Sonic is built for

Nova Sonic fits teams that already run on AWS and want a voice agent next to their existing services. Its feature list reads that way: expressive voices in seven languages, controllable turn-taking, voice and text in one session, asynchronous tool calls that keep the conversation going while a lookup runs, and a 256K context window.

It is a harder fit for:

• Indie and mid-size studios without an AWS account, IAM roles and a backend already in place.
• Unity and Unreal teams, since there is no engine SDK.
• Games with launch spikes, where many players talk at the same moment.
• Projects that need distinct character voices, beyond the preset voices.

Even for a studio already deep in AWS, the model is only one piece. Character design, voices, the engine client and the link to game state are still on your side of the line.

Four limits to check before you build

1. Concurrency. Bedrock's default quota for Nova 2 Sonic is 20 concurrent requests per account per region, and it is marked not adjustable in Service Quotas. AWS does accept increase requests through a support form, with no guarantee and priority for accounts already using their allocation.
2. Connection length. Each connection is limited to 8 minutes. AWS publishes a renewal and session continuation pattern in its samples, which you adapt and test.
3. Regions. Nova 2.5 Sonic runs in four regions: US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo). Players far from all four add round-trip latency.
4. Clients. AWS samples cover Python, Java, Node.js and .NET. There is no Unity sample, and AWS credentials must stay on your server, never in a game build.

A launch-day example

Picture a co-op game with a voiced quest giver in the hub area. At launch, 2,000 players are online in Europe, and one in fifty is talking to an NPC at any moment. That is 40 simultaneous conversations, double Nova 2 Sonic's default quota for the region. The other 20 players get an error, a silent NPC or a fallback line.

You can request a higher quota ahead of launch, but approval is not guaranteed, and the request competes with your other launch work. Meanwhile, each of those conversations reconnects every 8 minutes, through a server you run, with credentials you manage.

To cope, a game team typically builds:

• A conversation queue with fallback lines, so a waiting player hears something in character.
• Region routing, sending each player to the nearest of the four regions and spreading load across quotas.
• A connection renewer that swaps streams every few minutes without cutting the NPC off mid-sentence.
• A credential service that keeps AWS keys on your servers and gives each client only what it needs.

None of this is a flaw in Nova Sonic for its intended use. A support line can queue callers. A game cannot queue players who want to talk to the blacksmith.

Amazon Nova Sonic vs Neocortex

Getting a voiced NPC running on Neocortex

1. Create a free account. No cloud account or IAM roles needed.
2. Build the character in the visual editor: Agent, Personality, Emotion and Knowledge for who it is, Intents and Actions for what it can do.
3. Clone a voice from a short sample you have the rights to, or design one from a text description.
4. Test it in the browser. Editor testing never uses credits.
5. Install the Unity SDK and open a voice sample scene, or use Unreal Engine 5, Godot (beta) or the Web API.

From there, your game sends nearby objects and recent events, and the character replies with lines tagged by emotion and actions such as { "name": "OPEN_DOOR", "targetId": "door_02" }. On Pro and Team, up to six characters can share one conversation, so the whole hub can talk. Conversation Flows add branching story beats, and the API returns checkpoint keywords (flowState) when the player reaches one, so the quest giver can move the story forward instead of only answering questions.

The short version

Nova Sonic is a speech-to-speech model for customer service agents on AWS, with quotas, connection limits and credential handling designed for that world. For game NPCs, Neocortex gives you voiced characters, engine SDKs and gameplay hooks on a self-serve plan, so launch prep is about your game, not Bedrock quotas.

AWS details were last checked in October 2026.

Frequently Asked Questions