At a glance
Latitude builds AI-native games: experiences that are only possible because of AI, not games that use AI to speed up production. Its first title, AI Dungeon, pioneered open-ended AI storytelling and proved there was large-scale consumer demand for interactive narrative play, reaching millions of players and tens of thousands of creators.
Voyage, the studio's next-generation platform, is an AI-native RPG where nothing is scripted. Players describe worlds in plain English and adventure through them with friends, while a deterministic World Engine enforces real RPG mechanics on top of the AI: skill checks, HP, inventory, and persistent memory, so every choice carries consequences. Across its titles Latitude counts more than 8 million registered players and over 95 million adventures created, averaging above 110,000 a day.
The studio is profitable, with a team of just over 30 people across seven countries, and is backed by NFX, Google's AI Futures Fund, Griffin Gaming Partners, Album VC, Midjourney, and executives from Roblox.
The problem
Always-on voice at consumer scale
Voice in Voyage is not a garnish. Every story beat can be read aloud, and every NPC a player meets needs to sound like a distinct character rather than a narrator reading another line. That turns TTS from a feature into a load-bearing cost center: a game with narration on by default generates speech continuously, for every player, on every turn.
The incumbent voice provider powered AI Dungeon's original narration and set Latitude's quality reference, but at roughly four times the cost for this use case the pricing did not work for a consumer product where every beat is narrated and every NPC needs its own voice. What Latitude needed was studio-quality dynamic voices at consumer-scale unit economics, and the two rarely arrive together.
The team ran a full bake-off before settling. One alternative was evaluated end to end, but inconsistency between dynamically generated voices was the deal-breaker: a cast that shifts character between generations breaks the illusion the whole game depends on. Another was three to five times slower in Latitude's own production data. Self-hosting open-source models carried its own operational overhead. Inworld was the one option that held voice consistency across a dynamically generated cast while staying fast enough for narration that starts the moment text appears.
Inference had a parallel problem. Latitude's models ran primarily on a separate inference provider, where latency degraded as traffic scaled.
The solution
One partner for voice and inference
Latitude chose Inworld on three things, in that order: unit economics that make always-on voice viable in a consumer game, voice quality and consistency good enough for a dynamic cast where every NPC is assigned a persistent, distinct voice automatically, and how the team operates. Inworld prescaled endpoints ahead of Latitude's releases, worked directly with its engineers on voice markup and model behavior, and structured pricing so that going deeper improved the economics on both sides of the stack.
Serving voice and inference from one partner made those economics compound rather than stack. For a small team running an AI-native product at consumer scale, that flexibility mattered as much as the benchmarks did.
Integration took days. The voice work shipped as the launch narration provider for Voyage. On inference, Latitude moved its first production task in a day, half its engine traffic inside a week, and full story-model traffic within about a month.
Use case 01
Every narrator and every NPC
All eight Voyage narrator voices and the entire dynamic NPC voice cast run on Inworld TTS. Characters are assigned a persistent voice automatically, so an NPC a player meets in one session still sounds like themselves in the next, without anyone hand-casting a role. Roughly 97% of Voyage's TTS volume runs on Inworld, more than 15 million calls in six weeks.
Use case 02
World Engine tasks and story generation
Voyage runs roughly 20 AI tasks per turn alongside story generation. Inworld-hosted models serve the World Engine tasks that enforce skill checks, HP, inventory, and persistent memory, plus story and dialogue generation: more than 25 million calls in six weeks. Because the World Engine is deterministic, its outcomes have to land fast enough to feel like game mechanics rather than a model thinking out loud.
Use case 03
Moderation and Latitude's own fine-tuned models
Latitude routes content moderation through Inworld's chat completions endpoint, and Inworld hosts Latitude's own fine-tuned story models alongside the first-party ones, so the studio keeps the models it trained without running the serving stack itself. The team has also prototyped Inworld streaming speech-to-text for voice input.
Results
Narrating every story beat for every player costs about 6% of Voyage's total AI spend, roughly a quarter of what the incumbent premium vendor charged for comparable quality. In production, Inworld TTS averages 0.9 to 1.4 seconds per call against about 4.3 seconds for the closest alternative Latitude still runs, which is the difference between audio that begins as the story appears and audio the player waits for.
15M+
TTS calls in six weeks
25M+
Inference calls in six weeks
0.9-1.4s
Average TTS latency in production
6%
Of total AI spend to narrate every story beat
Voice became one of Voyage's signature features rather than an accessibility option, and narration now ships on by default. Community feedback on Inworld TTS has centered on the reading itself: punctuation understood, natural pauses, better cadence and pacing.
What’s next
Deeper on voice, and more of the stack
Voyage is in open beta today on the App Store, Google Play, and the web, with no waitlist and no invite codes, alongside DM Mode, public multiplayer servers, and Voyage Studio for instant world creation.
With Inworld, Latitude plans to expand its voice roster using voice design, complete its model traffic migration, scale Inworld's hosting of its fine-tuned models, and explore voice input and real-time voice in multiplayer.
Let's talk
Where this started, and where to find us
The partnership started at GDC in San Francisco, where Latitude's team sat down with Brian Cox from Inworld to talk through what always-on voice would take at consumer scale.
If you're building AI-native games, the Inworld team will be back at GDC. Contact us to set up a meeting.
