Cinema Is Sound

As a visual artist and film-maker, I saw AI coming for my job years ago.

Knew the light at the end of the tunnel was a train.

Everybody thinks AI can’t do their job.

Well. AI can have mine. Most of it.

I just want to keep the fun, soulful parts.

And audio mixing is not one of them.

It is right up there with removing the ‘umms’ and ‘ahh’ from hours of interviews with unprepared twats reading from marketing BS.

AI now removes the umms and pregnant pauses and mixes my stuff.

Hits my DaVinci timeline; checks all the levels. Sweetens. Dips. Fades. Friggs with the hertz. Magic.

Do I have a problem using AI? Not. One. Bit.

When it came to the sound design for Carny Island, I knew if I gave ElevenLabs my best character voice, I could use its studio to add age and an accent.

The actors I spoke to about the job wanted so much money the production would not have been possible. So I did it myself.

I didn’t know ElevenLabs also had a sound-effects generator.

I own Logic Pro but the software is so complex I can’t get past New Project.

CODEX knows it inside and out.

I had enough Carny Island lore to keep writing songs forever.

I didn’t know SUNO could hammer out hits.

So here’s the process; and yes, AI summarised this from three days of an intensive session on CODEX Sol 5.6 at Medium. I use it to check spelling, clean up my tenses and work out whether I should use its or it’s.

It’s an exchange between the game’s host and the player who’s about to rob a grave—but the player is bleeding and blood on Carny rings the dinner bell.

Most studio executives want you to deliver the whole game these days. They’re all too busy lunching to realise they don’t belong to a studio anymore...

It was produced by one person, on a Mac laptop, using a compact professional toolchain I already owned and subscribed to.

This was never a $200 studio. Logic Pro and DaVinci Resolve Studio alone cost A$768.99 at their current Australian prices. The recurring stack also includes Codex at $300, Claude Pro, ElevenLabs Pro and Suno—before the Mac, headphones, storage, creative labour or the years spent developing Carny Island.

The honest comparison is with the conventional production that this small system replaced. Using conservative Australian award minimums and published freelance rates, the equivalent voice, music, Foley, editing and mixing work for this pilot would cost approximately A$8,000–A$12,300. That excludes composer fees, usage and licensing, further revisions and the larger Carny Island sound library.

The important point is not that AI made the soundtrack cheaply.

It didn’t.

A human made it, using AI alongside Logic Pro, in much the same way that a small studio uses microphones, samplers, performers, sound libraries and processing equipment.

Suno produced musical performances and stems from human-written direction.

ElevenLabs produced voice performances and raw sound-effect material.

DaVinci Resolve prepared the original recordings used to build the Uncle Shrift voice.

Logic Pro became the recording studio, editing room and mixing desk.

Unity will control when those sounds play, how they are randomised and where they sit inside the game.

None of these tools could produce the soundtrack by itself.

Suno did not know where the music belonged.

ElevenLabs did not understand the character, the scene, the animation or the timing.

Logic Pro did not decide which metallic strike sounded like a sledgehammer hitting rock, or whether the Villain’s grunt sounded too human.

Those decisions were made by listening, rejecting, rewriting, retiming and rebuilding.

That is the work.

The original material was cleaned conservatively. We removed only the large, unusable sections of silence. Natural pauses, changes of delivery and clean retakes were retained because they contained useful information about the voice.

The final training master ran for approximately 56 minutes and 39 seconds. It was exported at 24-bit, 48 kHz, with a peak of −3.5 dBFS and no clipped samples.

That master was used to build the Uncle Shrift clone in ElevenLabs.

The first generations were not automatically successful. Shrift needed to sound older, darker and further removed from my ordinary voice. French-accent prompts produced almost no detectable French accent. Russian produced a much stronger colour, but some versions moved too far away from the original performance. Reverb made the voice airy. Slowing it too much made it laboured. Poorly spaced scripts caused words to tumble over one another.

We tested, listened and adjusted.

The resulting voice uses the clone as its foundation, but the performance still depends on writing, punctuation, spacing, direction, model choice and human selection.

The clone does not play Shrift by itself.

It has to be directed.

It is a tiny, three-foot creature in a skeleton suit. It talks to itself, complains, schemes, grunts, giggles and becomes increasingly excited as it robs a grave. The audience should eventually be able to decode parts of its language.

Several directions failed.

The early “alien” noises did not work. Some grunts sounded like ordinary human exertion. Some giggles were too large. Other attempts sounded like puppies, pigs or generic cartoon creatures without becoming a character.

The useful direction emerged through repeated listening:

higher pitched;

faster;

slightly sing-song;

male and female performances overlaid;

both voices occupying nearly the same moment;

approximately five per cent faster;

small, spiteful and amused;

capable of moving from muttering to panic almost vertically.

The language became SCRITCH.

Words such as Vinni Gruk, Nur val and Kri-kri-kri were written for particular dramatic purposes. The Villain can talk about Vinnie’s grave, Uncle Shrift, Crookwood, its mother and its mission without simply speaking English.

That voice was not found in a preset.

It was developed through performance direction and editorial selection.

The player digs through six substrate layers. Each one needs a distinct material, tool and physical response.

Layer 6 is the final black-earth substrate above the coffin. It contains shovel scrapes, loose particles, earth landing above the pit and occasional metallic contact with the coffin lid. We created seven variations so Unity can randomise them. Some favour the shovel. Others end with a louder metallic ring.

Layer 5 is rock. The action lasts 33 animation frames at 12 frames per second, so the complete event has to fit inside roughly three seconds. It begins with three sharp, ringing sledgehammer strikes. The rock then breaks, and the remaining sounds become duller and heavier. The pitch and force of the blows vary so the sequence does not sound like the same sample repeated.

The Villain’s exertion was developed separately. Its grunts and muttering were added only after the physical tool sequence worked. When the Villain lifts an intact boulder out of frame, the rock does not mysteriously shatter. The boulder movement belongs at the end.

Layer 4 is clay. That was unexpectedly difficult. Mud sounded too wet. Ordinary digging sounds were too brittle. Some attempts resembled wind hitting a microphone. The useful texture came from treating the material more like dense dough or soft meat: yielding, resistant and smooth, with the grub hoe cutting and pulling through it.

Layer 3 is soft dirt and a shovel.

Layer 2 is a pick striking hard dirt: a dull metallic thud, accompanied by sparse Villain grizzling. Seven variations were produced to prevent mechanical repetition.

Layer 1 is a rake moving through grass. This is the first time the player hears the Villain clearly, so the occasional giggle has to introduce the character rather than merely fill the silence.

These sounds were generated as ingredients, not accepted as finished effects. Logic Pro was where they became usable.

It is complex software. It does not remove the need to understand timing, tracks, levels, envelopes, pitch, space or the difference between a source recording and a finished cue.

It gave us a place to:

line up takes against animation timing;

cut unwanted beginnings and endings;

layer separate physical components;

alter pitch and speed;

control the balance between tool, material and character;

soften harsh transients;

add small amounts of echo;

reduce excessive reverb;

build distinct variations from shared ingredients;

audition sounds in sequence;

export consistent game-ready files.

A good shovel strike might contain several sources: the initial blade movement, the contact, a metallic ring, displaced dirt and an environmental tail.

The AI could generate each ingredient.

Logic Pro allowed a human to decide precisely when each one occurred and how much of it remained.

That is the difference between generating a sound and doing sound design.

bottles breaking and ringing;

bones and meat being cut;

eyes and body parts being removed;

organs being pulled from a body;

body parts landing in a shopping trolley;

slot-machine rewards for recovered loot;

a vicious dog attack;

the Villain’s death squeal;

an AK-47 burst with a long environmental tail;

pneumatic interface movement;

button buzzes and UI feedback;

Uncle Shrift’s dialogue and death remarks.

The Ballad of Vinnie remains separate from these interactive sounds.

That separation matters. Music should be able to continue, change or duck while Unity triggers physical events independently. A shovel strike should occur because the shovel hits the ground—not because it was permanently baked into a music track.

The soundtrack is therefore a system, not one long audio file.

It would be equally dishonest to say AI made it.

I think AI will let the talent of real artists rise to the surface—and see the other 80 per cent produce the slop.

The characters, language, performances, dramatic timing, scene construction, rejections and final choices came from a human production process. AI expanded the number of raw performances we could audition without hiring a Foley stage, a cast, a recording engineer and a sound library for a pilot.

That reduced the cash cost.

It did not remove the creative work.

Most generated material was not accepted untouched. Sounds were rejected because they were too wet, too muffled, too repetitive, too human, too large, too slow, too airy or simply wrong for the picture.

The soundtrack improved through criticism.

The machine supplied possibilities. A human decided what the scene meant.

music generation;

voice cloning and performance generation;

character-voice development;

Foley and effects generation;

multitrack editing;

timing to animation;

layering and processing;

high-quality exports;

organised delivery into the game-production workflow.

A conservative conventional-production comparison is:

character voice actors, direction and retakes: A$1,800–A$3,600;

music performers, studio and mix: A$3,200–A$3,500 minimum;

Foley, sound-effects design, editing and layering: A$2,400–A$4,000;

voice-session engineering, deliverables and mastering: A$600–A$1,200.

That produces a working estimate of A$8,000–A$12,300 for the pilot workflow—not including a composer fee, usage rights, licensing or additional revisions.

The point is not that AI made sound for $200. The point is that one human used a compact, already-owned studio system to perform work that would conventionally require a small crew.

We also retained clean masters, created multiple variations and transferred the material with manifests and checksums rather than leaving a pile of anonymous downloads.

This is still a pilot.

The final mix must be tested inside Unity, against animation, music, dialogue, environmental sound and actual gameplay. Levels that work in Logic Pro may need to change once the player is moving through the scene. Spatial positioning, randomisation, ducking and runtime behaviour still belong to the game engine.

But the production system works.

One person can move from drawing to writing to coding, open a Mac laptop, launch Logic Pro and continue building the soundtrack without waiting for a conventional studio session.

That does not make the work automatic.

It makes the studio smaller.

And cinema is still sound.

Price and rate references checked 28 August 2026: Apple Logic Pro Australia; Blackmagic Design DaVinci Resolve Studio Australia; Fair Work Live Performance Award MA000081; Fair Work Broadcasting, Recorded Entertainment and Cinemas Award MA000091; MEAA Voiceover Industry Rate Card.