A conversation should belong inside the world
Voice is one of the clearest ways to make an AI companion feel present. Instead of opening a separate chatbot, a player holds a push-to-talk key, speaks naturally and hears the companion answer while the game continues around them.
The goal is not a theatrical voice demo that works once. The complete experience must survive ordinary microphones, accents, background noise, network changes, reconnects and real game state. That is why the voice update is being developed as an end-to-end gameplay path rather than an isolated speech feature.
What happens after you release push-to-talk
The launcher captures a short, bounded mono audio stream while the configured PTT key is held. Releasing the key finalizes one utterance. A public player gateway authenticates the device and game session before the speech service produces a final transcript for the current turn.
The transcript is treated as untrusted player input, not as a command with automatic authority. Java adds the authoritative character, party and world context and exposes only the semantic capabilities that are legal now. The AI planner may then propose speech and a bounded action. Java resolves internal IDs and validates the result again before anything changes in the game.
- Microphone audio is bounded by duration, bytes and chunk count
- Reconnect queues are bounded so a broken network cannot grow memory forever
- Sequence and idempotency rules prevent a retry from duplicating an action
- A player can disable voice and continue through text where supported
Speech is not authority
A perfect transcription can still describe an impossible action. A companion cannot create an item, cast an unknown skill, spend another player's currency or act on a stale target merely because the sentence sounds confident. The language model selects only from current opaque candidates, and the GameServer remains the sole execution authority.
This design also protects companion identity. A public IP address or a character name sent by the launcher is not enough. The production path binds a short-lived launcher session to the logged-in account and asks trusted server state which character is active.
Privacy and graceful failure
The intended default is transient raw audio processing, with no persistent microphone recording. Launcher logs record technical timing and error categories, not the audio, access tokens or full player speech. Any future transcript retention feature would require a separate purpose, consent and retention policy.
Voice must fail gracefully. If the microphone is unavailable, permission is denied or the speech provider is temporarily offline, the launcher should explain the exact problem and preserve the rest of the game experience. An AI outage must not become a reason to corrupt game state or trap the player in an update loop.
What the beta will measure
The upcoming test waves will measure release-to-final-transcript latency, time to first TTS audio, reconnect behavior, playback underruns and recognition quality across real player hardware. English, Russian and Lithuanian are important early validation cases, with additional languages evaluated through the same evidence-based process.
The result we want is simple to describe and difficult to fake: hold to speak, say what you mean, receive an in-character reply and see only actions that the actual game rules permit.
Step into The World of AI
Follow the official route for the current test status, client preparation and account access.
Explore AI companions