Deepgram Voice Agent API for Game NPCs vs Neocortex

Deepgram's Voice Agent API can voice a game NPC, but you pay for every minute a connection is open and you build the character and the engine client yourself. Neocortex bundles the same speech pipeline with character tools and engine SDKs, on monthly plans.

Deepgram is a speech recognition company. Its Voice Agent API runs speech-to-text, a language model and text-to-speech over one WebSocket connection, with function calling and a choice of providers at each step. Here is what that costs and involves in a game, and how the two compare.

Costs at a glance

Deepgram publishes per-minute rates "calculated based on websocket connection time". The pay-as-you-go column:

Growth pricing lowers each rate. Bringing your own LLM or TTS also means paying that provider separately.

Neocortex plans are monthly, with credits included:

Testing a character in the browser editor never uses credits.

Why connection time matters in a game

On a support line, a call starts, the agent talks, and the call ends. Connection time and conversation time are nearly the same thing.

Games are different. A player walks up to an innkeeper, asks two questions, then spends twenty minutes looting the cellar before coming back. If the NPC's connection stays open the whole time, all of it is billed.

Take an open-world game where 1,000 players each spend an hour near a talkative NPC but only ten minutes actually speaking to it:

• Connection left open: 60,000 minutes, about $4,500 at the Standard rate.
• Connection opened only for conversations: 10,000 minutes, about $750.

The second number is reachable, but only if your client opens and closes connections around each exchange, and reconnects fast enough that the player does not notice. That logic is yours to write and test. In practice it means:

• Opening the socket when the player enters talking range or presses a talk key, not when the scene loads.
• Closing it after a silence timeout, and deciding how long that timeout should be for each NPC.
• Restoring the conversation context on reconnect, so the innkeeper still remembers the question about the cellar key.
• Hiding the reconnect delay behind an animation or a short line, so the player never waits in silence.

Each of these is a small feature. Getting all four right across every NPC type is a sprint of its own.

What you build on top of the pipeline

Deepgram lets you configure each moving part: managed LLMs from OpenAI, Anthropic, Google and NVIDIA, or your own OpenAI-compatible endpoint; Deepgram's Flux and Aura voices or managed Cartesia; and bring-your-own TTS from OpenAI, ElevenLabs, Cartesia or Amazon Polly, with fallback between providers. For a game NPC, you still own:

• A Unity client. The SDKs are Python, JavaScript, C# and Go. There is no Unity package or sample, so microphone capture, playback and the socket lifecycle are yours.
• The character. Persona, knowledge, tone and guardrails live in prompts and config you maintain in code.
• Game state. Function calling gives you a hook, and turning calls into door openings, item handovers and quest updates is your work.
• Voices. A cloned character voice means a separate provider account, contract and bill.

If you already prototyped on Deepgram

Most of your prototype carries over. Your system prompt becomes the starting point for the character: background into the Agent node, facts and lore into Knowledge (as text or PDF and TXT files), tone into Personality and a starting Emotion. Your function definitions become Intents and Actions, so instead of parsing a function call, your game receives an action name and the ID of the object it targets.

On the client side, the custom socket code goes away. The Unity SDK, Unreal Engine 5 integration, Godot (beta) plugin or Web API replaces it. If you used a bring-your-own TTS voice you have the rights to, clone it from a short sample in Neocortex so the character sounds the same.

Deepgram Voice Agent vs Neocortex: a decision matrix

Our take

Deepgram Voice Agent is a speech pipeline designed around calls that start and end, which is why it bills by connection time and leaves the character to you. For a game, Neocortex gives you speech in and out, the language model, voices and the character in one account, with engine SDKs and gameplay actions, so you pay for conversations rather than open sockets and spend your time on the NPC instead of the plumbing.

Deepgram details were last checked in October 2026.

Frequently Asked Questions