Connect speech, an LLM and game actions to prototype one bounded realtime NPC.
PRODUCT OVERVIEW
Inworld AI is a runtime stack for realtime interactive agents. It can receive a player's microphone audio, turn it into text, send the conversation and character context to a chosen language model, then stream a generated voice back. Memory, knowledge, turn-taking, back-channel behavior and tool calls can sit in the same observable pipeline.
Unity and Unreal templates provide a shorter path to a talking character. You describe the role, voice, motivations and constraints, connect the Runtime, and start from a text or push-to-talk demo. The game still owns the character model, animation, navigation, quest facts, permissions, rewards and saved state. Inworld is not a button that generates a complete NPC or game system.
This makes a first voice NPC approachable even if you cannot build the entire STT–LLM–TTS stack yourself. Start with one guide, witness, companion or optional hint character in a small scene. Programming and engine judgment become more important when dialogue can trigger actions or serve many players, but they are not prerequisites for exploring the first bounded interaction.
The indexed Unity route starts at Unity 2022.3 LTS and recommends 6000.0.41 or newer; the Unreal route lists 5.4–5.7 and a C++ project. Both begin with an API key and sample character. Current engine documentation redirected to login when checked, so confirm package versions and release notes inside the account before changing a production project.
Stream microphone audio through speech recognition, a selected LLM and TTS in one session. Turn detection and streaming reduce plumbing, but real game latency still depends on model, network, context and client behavior.
O05O06Character context, knowledge, conversation history, long-term memory and voice direction can shape an answer. Longer context can also increase token use, and generated statements are not authoritative game facts.
O05O06O08The Runtime can request registered tools without breaking the audio session. Treat each request as a proposal: your game validates the action, arguments, permissions and resulting state change.
O02O05O06U04The indexed quickstarts describe a Unity package with text/push-to-talk demos and an Unreal plugin with Character/Chat templates. Shipping still needs version compatibility, a credential backend, logging, reconnect behavior and target-device checks.
O03O04O15O16INTERFACE & EXAMPLES
The TTS Playground example combines text input, model and voice selection, speed controls, and an audio preview.
A first bounded voice NPC, a small Unity or Unreal prototype and a transparent public-rate budget. This page does not establish current Chinese quality, end-to-end latency, long-session stability, production concurrency, a measured bill or an accepted-conversation success rate.
Beginner scope: One constrained voice NPC using an engine example or WebSocket template.
Agent scope: Create a voice/text session and receive output events.
Confidence: learning Medium · Agent Medium
These are independent editorial judgments, not an overall score or a claim of hands-on agent integration testing.
KEY FINDINGS
Inworld combines speech recognition, a selected LLM, streaming voice, memory and tools in one observable path. Engine templates can get a character speaking sooner. Character assets, navigation, quests, permission checks, state transitions and failure UX remain your game's work.
O02O03O04O05O15Historical developers, players, journalists and a 48-participant study found novelty, emotional engagement or freedom in bounded scenes. The same evidence includes recognition errors, off-topic or incorrect answers, interruptions and connection failures. Positive overall experience does not mean reliable quest completion.
U01U05U06U08U09U10Most detailed game reports describe the 2023–2024 Character Engine. Recent records include one unresolved Runtime debugging question and two non-game TTS uses. That is enough to justify a small prototype, not a current success rate, Chinese-quality claim or production-maturity label.
O02O14U10U11U12U13The Cygnus team added its own intent layer for concrete game behavior, while media and research cases still saw misunderstandings or invented information. Register narrow tools, validate their arguments and keep consequential state changes deterministic.
O05O06U04U08U10STT is billed by player-audio time, TTS by generated characters and the chosen LLM by input/output tokens. Plans also change rates and request capacity. A low nominal session cost does not include subscription cash, retries, backend hosting, growing context or the share of conversations players accept.
O07O08EDITORIAL VERDICT
Inworld is worth considering when voice conversation is part of the game experience and you want one runtime instead of wiring speech recognition, a model and voice output from scratch. It gives beginners a real place to start: one character with one bounded job in a small scene. Keep rewards, inventory, combat, quest completion and saved progress behind deterministic game logic. The current Agent Runtime is newer than most independent game evidence, so prototype before making it a core production dependency.
BETTER FIT
POORER FIT
WORKFLOW FIT
Choose a guide, witness, companion or hint character. Write what it knows, what it must not claim and which one low-risk tool it may request. Prepare role questions, out-of-scope questions and interruption cases before integrating a larger scene.
Use the current Unity text/push-to-talk demo or Unreal Chat template in a disposable project copy. Add subtitles, one no-side-effect action and per-stage logs before custom animation, memory or multiple characters.
Invite a small audience and record recognition, time to first audio, full response time, interruptions, tool acceptance, disconnects and credits. Use those results to estimate accepted-conversation cost and peak capacity.
RECOMMENDED WORKFLOW
Start from a testable brief, then move through prototyping, assets, audio, testing, and release with an explicit handoff and human check at every step.
SHORTEST RESPONSIBLE PATH
This is a low-risk starting workflow synthesized from the evidence.
Pick one bounded character role and write a short source-of-truth sheet: identity, known facts, forbidden claims, tone and one allowed tool request.
Create twenty prompts covering ordinary dialogue, role boundaries, repeated questions, interruptions, silence, Chinese or accent needs and an unauthorized action request.
Connect the current Unity or Unreal sample in a disposable project copy. Keep credentials on a backend, add subtitles and expose only one action with no irreversible side effect.
Run the prompts on the target network and device. Record transcription, first-audio time, full response, rule compliance, tool validation, reconnect behavior and actual credits.
Let a few people try the same bounded scene. Only widen access after defining an accepted conversation and calculating real cost, peak sessions, privacy notice and a scripted fallback.
PRICING & RIGHTS
Pricing and terms last checked: Sep 1, 2026
Illustrative public-rate arithmetic, not measured billing: one ten-minute session contains five minutes of player speech and five minutes of NPC speech (5,000 characters), ten turns, 12,000 input tokens and 1,500 output tokens using Gemini 2.5 Flash at $0.30 / $2.50 per million tokens. Exclude free allowances, retries, memory/RAG growth, moderation, tools, backend, tax and labor.
O07O08You pay $25 and receive $25 in monthly credits at Creator rates, with ten concurrent requests and an estimated 40 user sessions. Unused value follows the rollover rules; it is not a pay-only-for-what-you-used plan.
One session consumes about $0.06568: $0.00833 STT + $0.05000 TTS + $0.00735 LLM. This is a credit division under the stated assumptions; Inworld does not sell a separate one-dollar session pack.
One session consumes about $0.11568: the same STT and LLM assumptions, with $0.10000 of TTS. Voice choice changes cost before quality, retry or acceptance is considered.
These are whole-session theoretical ceilings under the same ten-minute workload, with no other account use. They are not 380 or 216 accepted conversations and do not override request or session capacity.
Assume 20 creator sessions plus ten visitor sessions. Creator cash paid remains $25 that month. The twenty-session iteration count is a planning assumption, not a first-pass success rate or a complete NPC development cost.
Creator rates consume about $6.57–$11.57, but cash paid is still $25; On-Demand uses the displayed rates and excludes the unclear free allowance. This is 100 total visits, not 100 simultaneous sessions.
Nominal usage is not accepted output. The real denominator is conversations that meet your character, task, voice and player-experience criteria. Actual cost must include failed and repeated sessions, growing context, external services, backend hosting, engineering and operations. Peak capacity is separate: Creator lists an estimated 40 concurrent user sessions, while 100 simultaneous sessions would require a higher published tier or a confirmed custom arrangement and load test.
The pricing page offers a commercial-license path, and the general terms assign Inworld's rights in compliant outputs to the customer. The same public materials also mention internal-business use, deletion of Services, Models and Outputs after termination, and an older SDK agreement that ties Project Files to the Inworld Platform. Confirm current Runtime redistribution, post-termination game content, third-party model/voice rights and any cached audio before release. This review is not legal advice.
Player speech and chat can enter a data path involving audio, possible voice biometrics, research/training and third-party models. Do not assume an ordinary self-serve workspace has zero retention. Confirm player notice or consent, region and age requirements, DPA/ZDR settings, deletion, model-provider handling and the checkout terms that apply to the actual account.
PRODUCTION RISKS
The public product line has moved from Character Engine toward Agent Runtime and general realtime services. WebSocket may be GA while the overall Realtime API, Router or STT carry preview labels. Confirm the exact engine package, changelog and supported target instead of treating one GA label as maturity for the whole stack.
O02O03O04O14O16Historical evidence repeatedly includes recognition problems, waits, speaking over characters and sessions that stop responding. Current incidence is unknown. Add subtitles, interruption handling, timeout copy, retry limits, scripted dialogue and an exit path before testing immersion.
U05U06U08U09U10U13A persuasive character can still be wrong. Keep quest truth, inventory, rewards, combat, payments and saves in deterministic systems. Tool calls need narrow permissions, schema validation, idempotency and failure rollback.
O05O06U04U08U10Do not ship a reusable Base64 key in the client. Use a backend to keep credentials and issue short-lived tokens, then document what text, audio, metadata and logs leave the game, how long they remain and which model providers receive them.
O10O11O12O15O16U01U14The commercial-license label, general service terms and older SDK agreement do not answer every current Runtime distribution question in the same words. Before release, confirm plugin redistribution, cached outputs, post-termination use, attribution and third-party voices or models for the exact plan.
O07O09O10O13NOT VERIFIED
RESEARCH METHOD
Fourteen independent records across ten platforms and five evidence types: developer/community cases, player/review records, two independent media hands-ons, an academic study and a technical repository. Steam contributes 3/14 (21.4%) and Reddit 2/14 (14.3%); no platform reaches half the set. Searches also covered Epic Developer Community, GameDev.net, G2 and Hacker News, where results were absent, too weak, incentivized, derivative or inaccessible. Vendor stories and sponsored media were excluded from independent counts, while repeated demos, one game and the study's earlier thesis were clustered or deduplicated. The recurring historical themes are useful for workflow and failure-mode guidance, but recent exact-version game evidence is insufficient; current Runtime performance remains low-confidence.
Current product entry point for realtime speech, model routing and agent infrastructure. Vendor performance and ranking claims are not treated as independent results.
Runtime graph, templates, observability and the Unreal release. Unity was described as early access at publication, not as a current status guarantee.
Search-indexed Unity requirements, package installation, API key and text/push-to-talk demos. Direct access redirected to login when checked.
Search-indexed Unreal 5.4–5.7, C++ project, plugin, Character/Chat template and trace requirements. Direct access redirected to login when checked.
Public index for current TTS, STT, Router, Realtime API, memory, back-channel and tool documentation.
Single-session speech pipeline, model choice, tool calls and transport status. Published latency numbers remain vendor claims, not game measurements by this site.
Current self-serve monthly prices, credits, TTS/STT rates, request limits, estimated sessions, rollover and commercial-license label.
Current model list and input/output token rates. The page budget uses Gemini 2.5 Flash as a transparent example, not a quality recommendation.
Service license, inputs and outputs, material use, training exceptions, restrictions, fees and termination requirements.
Additional voice-input, voice-output, user voice model, content-rights and prohibited-health-data terms.
Chat, audio, possible voice biometrics, research/training, retention and cross-border processing. Enterprise processor data has a separate scope.
SOC 2 Type II, encryption and Trust Center entry points. Zero data retention is not presented as a default on every self-serve plan.
Older public SDK agreement covering Project Files, distribution and platform restrictions. Its applicability to the current Agent Runtime needs confirmation.
Official status labels distinguish TTS GA, Router/STT research preview, Realtime API research preview and WebSocket GA.
Search-indexed guidance not to expose Base64 credentials in clients and to issue short-lived JWTs from a backend.
Archived official sample documenting a session backend, WebSocket, retry, timeout and conservative concurrency handling. Archival does not prove the Runtime ended.
An independent JavaScript narrative-game integration needed a Node token service. The author found basic connection manageable and characters engaging, while warning that open chat can distract from the game.
An open-world story developer liked the early character setup but questioned whether its limits could support a story that evolves with player interaction.
A developer connected both Inworld and Convai but described the process as challenging and highlighted the STT–LLM–TTS stack and recurring connection costs.
The game team combined Inworld speech with its own intent layer for gathering, combat, skills and events. It explicitly separated generated conversation from executable NPC behavior.
One player found the official Origins demo compelling, then reported that all NPCs stopped responding and a restart lost progress. This is an older demo case.
A second Origins player reported NPC responses stopping after roughly ten minutes and recovering only after restart. It is a reproduction clue, not a current Runtime test.
A 29.2-hour player disliked the game's AI NPC integration and noted that it could be disabled. The review also criticizes other systems, so not all dissatisfaction can be assigned to Inworld.
A journalist found one effective on-path exchange, while also encountering multi-second waits, characters speaking over each other and mixed off-topic responses. The demo was too rough to prove replacement of written dialogue.
A Japanese reporter compared an earlier accent failure with a newer live demo that still looped or misunderstood topics but felt more continuous and emotionally engaging.
A 48-participant VR study published in 2026 used a 2023 Inworld build. Overall experience was positive, while 77% reported recognition issues and notable minorities reported unhelpful or incorrect answers.
A product developer reported choosing Inworld voice for Korean narrative users. This is a current, low-weight TTS clue, not evidence for the game Runtime.
A developer used Inworld voice in a calling agent and described it as more conversational. The non-game case cannot establish NPC engine quality.
A current Runtime user could not identify which pipeline stage caused delay. Staff pointed to per-node debug logs, but the thread records no final measured resolution.
An independent open-source scene includes a remote server, proxy/WebSocket, custom chat UI and environment endpoints. It shows integration glue, not output quality, and uses an older API path.
DISCLOSURE
Owner Review passed on September 1, 2026, accepting this bounded public-evidence review and its medium-confidence conclusion. This work introduced no affiliate link, sponsorship, vendor account, free credits, equipment, interview or technical support, and no Inworld SDK or model was tested. The owner's prior use and commercial relationships remain unstated; approval does not resolve the listed Runtime, pricing, distribution or data unknowns.
This public research review will be revisited when pricing, terms, product versions, or material new evidence changes.
