How Chimera, Jameson Hodge’s Physics-Based Vocal Synth, Gives the Artemis VII Coach a Real Voice

Jump to comments

Chimera's 3D wireframe head on chimera.audio, a glowing blue see-through human head with the tongue, jaw, lips and airway drawn in green, red and pink as the synth speaks
Chimera draws its simulated vocal tract as a see-through 3D head, with the tongue, jaw and lips moving as the voice speaks.
Screenshot: Chimera.audio by Jameson Hodge, 2026

Artemis VII, the free starship-catching game on the Aglena Games page, has a flight coach that sits in the corner of the screen as a glowing wireframe head. When you come in too fast, it says "You’re coming in too hot!" When you crash into a tower, it gets sarcastic. None of those lines were ever recorded. There is no voice actor, no sound file and no text-to-speech service behind them. Each one is built from scratch, in your browser tab, by a simulated set of lungs, vocal folds, throat, tongue and lips.

That simulation is Chimera, an open-source vocal synthesizer by developer Jameson Hodge. It went viral on X in early November 2026, and it is one of the most impressive pieces of audio software released this year. This is how it works, and how Artemis VII turned it into a coach with moods.

The short answer: Chimera models the human airway as a branching network of acoustic tubes and computes, sample by sample, how air pressure and airflow move through it. The engine is written in Rust and compiled to WebAssembly, so it runs in real time inside a web page. Artemis VII builds a copy of that engine into the game file, turns each coach line into mouth and throat movements, and sends them to Chimera on the browser’s real-time audio thread. The same engine reports the shape of the tongue, lips and vocal folds about 60 times a second, and the game draws the talking head from that data.

Turn the sound on: the Artemis VII coach speaks six real game lines in its calm, alarm, cheer and snark moods, all generated live by Chimera.
Gameplay capture: Aglena Games, 2026 (coach panel enlarged, live audio)

Meet Chimera

Chimera’s own description is short: "A physical-modeling voice synthesizer in the browser. Every sound comes from a simulated airway (lungs, vocal folds or syrinx, a branched vocal tract, lips, nose), not from samples." Its GitHub repository was created on Nov. 5, 2026. By Nov. 21 it had about 2,900 stars and 365 forks, and it is released under the MIT License, so anyone can use the code, including in a game.

The day after the repository went up, Hodge posted a 39-second demo with the deadpan caption "I only make serious software." The clip shows a cutaway head singing, with the airway lighting up as the sound moves through it. As of Nov. 21 the post had been seen more than 5.1 million times and liked more than 48,000 times.

Jameson Hodge’s Nov. 6, 2026 demo of Chimera, the post that made it go viral.
Video: Jameson Hodge, via X

Open chimera.audio and the instrument is right there. There are five modes. Play lets you drag the tongue, lips and jaw in a live cutaway of the head. Speak turns English text or phonetic spelling into speech. Sing sets typed lyrics to a tune or to the keys you play, including from a MIDI keyboard. Beatbox builds every drum sound from mouth movements. Bestiary swaps the human for a songbird, a seal, an alligator or an elephant. The body blends continuously between male, female and child.

The Chimera web app in Sing mode, with a cutaway of a human head and vocal tract on the left and lyrics, pitch, spectrum, voice dials and a keyboard on the right
Chimera’s Sing mode shows the simulated vocal tract beside lyrics, a pitch roll, the live spectrum and the keys.
Screenshot: chimera.audio by Jameson Hodge, 2026

It is also more than a web page. A command-line tool renders speech, songs and beats to WAV files, and a plugin version runs inside music software as a CLAP, VST3 or AU instrument on Mac and as CLAP or VST3 on Linux. The project’s plugin page says a Windows build isn’t released yet.

How a simulated airway makes a voice

Most synthetic voices start from recordings. Vocaloid-style singers stitch together short clips of a real person, and modern text-to-speech models learn from hours of recorded speech. Chimera skips the recordings entirely and simulates the physics instead. Hodge put the difference plainly in a follow-up post:

My vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings.

Jameson Hodge, on X, Nov. 7, 2026

The project’s architecture notes go into detail, and they read like a short course in speech science. Three ideas do most of the work.

A source: lungs and vocal folds

Lung pressure pushes air up the windpipe. At the larynx, a modeled glottis opens and closes in pulses that set the pitch. The model adds the small imperfections real voices have, such as jitter, shimmer, breathiness and vocal fry, so the voice doesn’t sound like a pure buzz. Birds get a different source, a two-sided syrinx built from a published model of birdsong.

A filter: a branching tube network

Above the larynx, the throat, mouth and nasal cavity are a connected network of short tube sections, each with its own cross-section. Chimera simulates pressure and airflow through that network step by step, a method known as finite-difference time-domain (FDTD) modeling, running internally at 96,000 steps a second for an adult human voice. The notes explain why this was chosen over the simpler textbook tube model: it handles branches such as the nose, air building up behind a closed mouth, the burst when a "p" or "t" releases, and walls that give a little. Hiss in sounds like "s" and "sh" comes from the simulated airflow turning turbulent at a narrow gap, not from a noise sample.

Movement: gestures instead of sounds

Words are not stored as sounds. English text is looked up in the CMU Pronouncing Dictionary, with a letter-to-sound fallback, and turned into a gestural score: a timeline of overlapping movements for the lips, tongue tip, tongue body, soft palate and glottis. Each articulator moves at a realistic speed, with the tongue body still gliding toward the next vowel while the lips close for a consonant, as real mouths do. The speech you hear is the result of those movements.

Chimera's Speak mode saying Hello, can you hear me, with a gestural score for lips, tongue, soft palate and glottis, a consonant chart, a tube diagram and a vocal-fold view
In Speak mode, Chimera shows the gestural score it plays for each phrase, plus the airway tube and the vocal folds in motion.
Screenshot: chimera.audio by Jameson Hodge, 2026

This is why the result sounds unlike anything else. It isn’t flawless. Hodge says he is still tuning it against recordings of real singers to improve how natural it sounds, and the project’s own docs list open items. But it is alive in a way sample-based voices are not. Change the shape of the mouth and the sound changes the way a real voice would.

How Artemis VII puts Chimera in a game

Artemis VII asks you to fly a pixel starship between launch-tower "chopsticks," landing pads and bare ground across Earth, the Moon, Mars and Titan, over 16 legs. Our full Artemis VII review covers the gameplay. The coach, called Capcom on most legs, gets its voice from Chimera. The game’s start screen credits it: "Voice: Chimera vocal synth · github.com/jamesondh/chimera-vocal-synth."

Getting a research-grade vocal tract to run inside a single-file browser game took a few deliberate steps:

  1. Pin and build the engine. The game vendors Chimera’s source at one upstream commit and compiles the Rust engine to WebAssembly with Cargo, using the project’s own chimera-wasm crate.
  2. Inline everything. The WebAssembly module, the audio-worklet code and the pronunciation dictionary are packed into the game file itself, so the coach never has to fetch its voice from the network.
  3. Plan on the main thread. When the game wants to say something, Chimera’s English front end turns the text into phrases, and its planner turns those into gestures, with pitch, speed and emphasis set by the coach’s mood.
  4. Speak on the audio thread. The game sends three messages to Chimera’s AudioWorklet: the body (a male or female vocal tract), the voice source settings, and the gesture schedule. The worklet renders the speech in real time and feeds it into the game’s own audio mix, where the synthwave music ducks under the voice.
  5. Let it rest. A small local patch adds a "sleep" message, so the worklet idles between lines instead of rendering silence. Another patch adds a fallback for browsers that can’t hand a compiled module to a worklet.
  6. Draw from telemetry. The worklet streams back the positions of the articulators, the airway pressures and the vocal-fold motion. The 3D head in the corner is drawn from that live data, so its mouth moves because the simulated mouth is moving.
Diagram of how Artemis VII's coach talks: game logic, text to phonemes and a gesture planner on the main thread send messages to the Chimera WebAssembly engine in an AudioWorklet, which returns audio and telemetry for the head panel
How a coach line travels from the game’s logic through Chimera’s planner and WebAssembly engine to your speakers and the animated head.
Diagram: Aglena, 2026, based on Chimera’s documentation and the Artemis VII coach code

There is one honest limit. The browser only allows an AudioWorklet on a secure (https) page, so the coach speaks on aglena.com but falls back to on-screen captions where the page isn’t secure. Every line still waits as long as it would take to say, so the timing of the game doesn’t change.

Four moods from one physical voice

Because Chimera is a physical model, a mood is not a different recording. It is a different way of using the same body. Artemis VII defines four moods, and each one changes the planner’s pitch, speed and emphasis, plus the voice source’s lung pressure, vocal-fold tension, breathiness, vocal fry, roughness and brightness.

  • Calm sits around 104 Hz at a slightly slow pace, for leg briefings and steady advice ("Ease to the left.").
  • Alarm rises to about 122 Hz, speeds up and pushes harder, for warnings like "Pull up! Pull up!" and "We’re out of fuel!"
  • Cheer sits near 118 Hz with the most emphasis, for "Yes! The arms have her! Nice catch!"
  • Snark drops to about 100 Hz, slows down and adds a hint of creak, for lines like "Oh, great. Another crater."

We checked this on the live game. Running Artemis VII in a headless Chrome browser and reading the engine’s own pitch telemetry while the coach shouted an alarm line, we measured about 117 Hz on its voiced sounds, well above the calm and snark settings. The coach’s jokes come at most every 15 seconds and never cut off a warning, so the voice stays useful.

Four frames of the Artemis VII wireframe coach head mid-sentence, with the tongue, lips and jaw in different positions
Four moments from one alarm line, with the tongue, lips and jaw drawn from Chimera’s live telemetry.
Screenshots: Aglena Games, 2026

The last leg, the Tile Job, goes further. Two foremen, Mack and Rae, show up to boss you through a heat-shield repair, and you pick one. Rae uses Chimera’s female vocal tract, with her pitch raised to suit the shorter tract. The game’s code notes why the multiplier stops where it does: pushed past about 220 Hz, "the words thin out." That kind of tuning is only possible because the voice is a model you can steer, not a fixed recording.

Why the engine can run in a game at all

Simulating air through a tube network tens of thousands of times a second is heavy work, and a game also has to draw frames and run physics. Chimera can share the space because of how Hodge built the engine.

The project started in TypeScript and was ported to Rust. The project’s engine design notes report that rendering a male voice offline took 46% of one CPU core in TypeScript in Chromium and 32% as WebAssembly, and that in the running app the human voice’s load fell from 40% to 29%. The port was checked against the old engine sample for sample, and the native and WebAssembly builds render bit-identical audio on x86-64.

Just as important for a game, the engine follows a strict real-time contract. Steady-state rendering allocates no memory, the gesture score has a fixed size, and output does not depend on the audio block size the browser picks. Those rules are what keep a voice from crackling or stalling while the game is busy. They were written with Hodge’s own longer-term plans in mind, but they are exactly what a game needs too.

What other people are building with it

Artemis VII is one of a wave of projects. Within days of the first demo, people were remixing it. AI podcaster Gavin Purcell gave the repository to an AI coding agent and got back a pair of singing cowboys. Developer Roman Pshichenko built a browser game where you play DOOM inside Chimera’s moving airway while it sings the E1M1 riff. The AI For Humans podcast gave Chimera a segment in its Nov. 13 episode.

Hodge has kept making music with it himself. On Nov. 9 he released "Just You," a song he called "a love letter to Astrud Gilberto and Walter Wanderley," with vocals by Chimera.

“Just You,” a song Hodge released on Nov. 9, 2026 with vocals by Chimera.
Video: Jameson Hodge, via X

What we know and what’s still open

What we know

Chimera is free, open source under the MIT License, and runs in the browser through Rust and WebAssembly. It speaks English, sings lyrics in English, Korean or Russian, beatboxes, and voices four animals. Artemis VII uses it for every spoken coach and foreman line, entirely on the player’s device.

What’s still unclear

How natural Chimera can eventually sound is an open question, and Hodge describes it as ongoing work. The project’s design notes list pieces not built yet, including a self-oscillating vocal-fold model and a wider bestiary, and bit-identical output on ARM-based Macs is still untested. Artemis VII also pins a specific Chimera version, so new improvements reach the game only when it is rebuilt.

Try it yourself

Bottom line

Chimera is the rare piece of software that is both serious science and pure fun. Jameson Hodge built a real-time physical model of the human voice, made it fast enough to run in a browser tab, gave it away under an open license, and wrapped it in an interface that makes people want to play. That is what let a small browser game give its coach a voice of its own. When the Artemis VII coach tells you you’re coming in too hot, you are hearing air move through a simulated throat, and that is a remarkable thing to hear.

Further reading

Cover of Elements of Acoustic Phonetics by Peter Ladefoged
Elements of Acoustic Phonetics

Peter Ladefoged’s short, classic primer on how the vocal tract shapes sound, the science Chimera turns into code.

Cover of Designing Sound by Andy Farnell
Designing Sound (MIT Press)

Andy Farnell’s guide to procedural audio, building game and film sounds from models instead of recordings.

Cover of The Rust Programming Language, 2nd Edition
The Rust Programming Language, 2nd Edition

The official Rust book, the language Chimera’s engine is written in before it is compiled to WebAssembly.

Audio-Technica ATH-M50x studio headphones in black
Audio-Technica ATH-M50x Professional Studio Monitor Headphones

Closed-back studio headphones for hearing the breath, fry and consonant bursts a physical voice model produces.

Cover of This Is Your Brain on Music by Daniel J. Levitin
This Is Your Brain on Music

Daniel Levitin’s bestseller on how we hear pitch, timbre and rhythm, a friendly companion to playing with Chimera.

Aglena Best Of product picks

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top
Aglena