
Local Asset Forge
I am testing local AI for art and audio, from stylized props in Unity to NPC voices made from text descriptions. The latest video explores sound effects, ambience and instrumental music through prompts, results and practical edits.
Project filmUsing Local AI to Generate Sound Effects and MusicWatch on YouTube ↗Using Local AI to Generate Sound Effects and Music
A creaking wooden lid, birds in a forest, and a rock track all start with text prompts. This video explores what MOSS SoundEffect can produce for effects and ambience, and what ACE-Step can make with different instruments, tempos and styles.
Watch the full video on YouTube
Hear acoustic folk, heavy rock, lo-fi, and fiddle/concertina music alongside the sound-effect auditions. I show the prompts, a replay of recorded events from a local music run, and the edits between a generated file and something I can use: trimming a late impact and blending a section into a loop. I still need to choose takes, check the timing, and listen for repetition.
Using Local AI to Make Voices for Games
A trader needs a greeting. A guard needs a warning. I used VoxCPM2 through VoiceStudio to turn text descriptions into eight NPC voice auditions, then used one generated voice as the reference for more dialogue. My Mac handles the writing and review; an RTX 4080 laptop with 12 GB of VRAM generates the audio locally.
Watch the video on YouTube · Download the prompts and nine raw WAVs (4.3 MiB ZIP) · Copy the prompt sheet
The eight designs start without a reference recording. Each reads the same warning followed by a line for that character. Play a sample, then expand its description to see the exact prompt and dialogue. These are the original generated WAVs; the video’s auditions have level changes only.
Frontier trader
Seed 411Exact voice prompt and dialogue
Voice description
A man in his fifties. Deep gravelly baritone, tired American frontier drawl, dry humor. Close conversational delivery, clear words.
Dialogue
Keep your voice down. Something is moving beyond the gate. If you break it, you bought it.
Veteran guard
Seed 412Exact voice prompt and dialogue
Voice description
An adult woman. Low firm alto, clipped British English, calm authority. Measured military delivery, restrained urgency, clear words.
Dialogue
Keep your voice down. Something is moving beyond the gate. Stay behind me and watch the rooftops.
Goblin scavenger
Seed 413Exact voice prompt and dialogue
Voice description
A small male goblin. High nasal voice with a scratchy edge, playful Irish lilt. Mischievous and conspiratorial, clear words, no shouting.
Dialogue
Keep your voice down. Something is moving beyond the gate. Shiny things first. Heroics can wait.
Field medic
Seed 414Exact voice prompt and dialogue
Voice description
An adult woman. Warm mid-range voice, soft American accent, gentle and reassuring. Quiet concern, steady natural pace, clear words.
Dialogue
Keep your voice down. Something is moving beyond the gate. Hold still. I can stop the bleeding.
Ship computer
Seed 415Exact voice prompt and dialogue
Voice description
An androgynous synthetic ship computer. Smooth even mid-range voice, precise neutral English. Detached, measured delivery, minimal emotion, clear words.
Dialogue
Keep your voice down. Something is moving beyond the gate. Warning. The outer airlock is now open.
Village elder
Seed 416Exact voice prompt and dialogue
Voice description
An elderly woman. Airy slightly trembling voice, warm low register, gentle English accent. Patient storytelling rhythm, quietly serious, clear words.
Dialogue
Keep your voice down. Something is moving beyond the gate. That bell has not rung in years.
Cult leader
Seed 417Exact voice prompt and dialogue
Voice description
An adult man. Dark resonant bass, polished British accent, intimate and unsettling. Slow controlled delivery, a faint smile, clear words.
Dialogue
Keep your voice down. Something is moving beyond the gate. You came here looking for an answer.
Nervous scout
Seed 418Exact voice prompt and dialogue
Voice description
A young adult man. Light mid-range voice, natural American accent. Breathless nervous energy, hesitant but intelligible, conversational volume.
Dialogue
Keep your voice down. Something is moving beyond the gate. I counted three. Then the lights went out.
Keep the voice across new lines
An NPC needs several lines that sound like the same person. I use the first generated clip as a reference, with an exact transcript of what it says, then supply new dialogue. Reusing a description alone can produce a different voice.
Here is the frontier trader again. This request uses the first trader WAV and its full dialogue as the reference, with no new style description.
Frontier trader — new dialogue
Seed 419Exact voice prompt and dialogue
Voice description
Reference audio: the original frontier trader sample above. Reference transcript: its exact dialogue. No new voice description.
Dialogue
The road is closed until morning. Bring a lantern if you plan on coming back.
Marlow, the trader in Rustwake, a separate Duskbound fork, is the game example in the video. His original description was: “A deep, rough-voiced man in his fifties with a heavy smoker’s rasp and a tired frontier drawl, dry humor.” I chose one of three design takes and used that clip as the reference for his other lines.
One line, one audio file
I save each line of dialogue in its own audio file and import those files into Unity. The game triggers a stored clip when the character speaks. A greeting, a warning, or a reply can each be replaced independently; the speech model runs while making the files, rather than during gameplay.
The eight new designs are auditions, separate from the existing Duskbound footage. Before choosing a voice for a game, I would check pronunciation, consistency across several lines, and whether the dialogue comes through with the game’s music and effects playing.
Model, settings and commercial use
These samples use VoxCPM2, whose official model card explicitly permits commercial use under Apache 2.0. VoiceStudio’s app license also permits commercial and internal use, while its supported models have their own terms. The default OmniVoice weights have a noncommercial restriction; these runs explicitly selected VoxCPM2.
The run used VoiceStudio 0.5.6, 10 steps, guidance 2, speed 1, and seeds 411–419. The WAVs are 48 kHz mono. Denoising, output postprocessing, pronunciation rewriting and automatic chunking were off. The ZIP contains the exact model revision, settings and hashes. Saving a seed documents a take; results can still vary between runtimes.
The video narration is a local clone of my own voice. The Design screen is a still demonstrating the two inputs; the full descriptions and dialogue for these auditions were sent through VoiceStudio’s local API.
Why I built it
While making a small game, I kept needing another model, sound, or piece of music. I wanted to see how far I could get making those things on hardware I already owned. The question was whether I’d actually use the results in a game, not just enjoy watching them appear in a browser.
From an image to a game asset
The process starts with a concept image. I cut out the subject, generate a textured 3D model, clean it up in Blender, and bring it into Unity. I work from my Mac; an RTX 4080 laptop in another room handles the heavier generation. I built a small Studio interface to keep the inputs, settings, results, and failures together.
The source image matters more than I expected. Some of my best 3D results started with references made outside the local pipeline. Local image generation produced promising concepts too, but it also gave me glossy finishes, odd proportions, and details facing the wrong way. Those problems often carried into the 3D model.
What survived Unity
Props gave me the most encouraging results. The chest, shield, and sword were worth testing in Relic Siege, my rough wave-fighting prototype. A simple canister looked surprisingly good, though it needed a more faceted finish to sit with the other art. Even then, a close-up could look much better than the same asset from the game’s isometric camera.


Where it fell short
Reducing a model to a practical size can erase the details that made its preview exciting. That happened with a skeleton. The equipped character was another useful reality check: generated armor, helmet, sword, and shield could make an interesting silhouette on a pre-rigged base, but the grip and shield placement did not look natural. I would not call it a usable character pipeline yet.
More than 3D
I also tested local sound effects, music, and voice generation. They made the pipeline more useful for a prototype, and the video narration itself was generated locally. The Studio app ties the stages together, but it does not remove the need to select, edit, and test each result.
Where I landed
Selected stylized props are good enough that I would try them in a prototype. I would not expect to press a button and get finished game art. Source images, mesh reduction, and fit with the game still take judgment and cleanup, and characters remain the clearest gap. The original 3D pass is paused, with the raw runs kept for future work. The voice, sound-effect and music workflows above continue the audio side of the experiment.
Project timeline
Along the way
- Latest update
Local AI sound effects and music
Hear MOSS SoundEffect effects and ambience alongside ACE-Step instrumental music, from acoustic folk and rock to lo-fi and fiddle/concertina. The video follows the prompts and results, shows a recorded local music run, and demonstrates trimming a late impact and editing a loop.
Video update · Watch on YouTubeUsing Local AI to Generate Sound Effects and Music Local AI voices for game characters
I designed eight NPC voices from text descriptions with VoxCPM2, then reused one generated voice for another line. The video shares the exact prompts and the workflow behind Marlow, Rustwake's trader. Play the raw samples and download the prompts on this page.
Video update · Watch on YouTubeUsing Local AI to Make Voices for GamesThe result, on video
I put together a short film showing the real generation sessions, the models in Unity, and what survived the move from preview to game. The honest result is encouraging for selected stylized props, but it is not a one-click source of finished game art. I am pausing the experiment here with the raw runs and failures saved.
Video update · Watch on YouTubeI Built My Own Local AI Asset Forge — Is It Good Enough for a Game?The character test set the limit
A detailed skeleton lost much of its appeal in a smaller export. I then tried locally generated armor, helmet, sword, and shield on a pre-rigged character. The silhouette worked at a distance, but close views showed a loose sword grip and a shield that did not sit naturally on the arm. That is not a repeatable character pipeline yet.
Testing a lighter sword
A revised local sword — one guard gem, a clean cutout, and a lighter blade — made it through reconstruction, export, and an equal-height comparison in Unity. The blade reads better than the darker version, and a test build rendered it held and registered a melee hit. It is a candidate, not a replacement. The sword already in the game stays until facing, alignment, and performance check out.
A better local image model
FLUX.2 Klein produced a better local chest than my earlier image model, and its export got a positive review when I looked at it in the game. A local shield made it through the same path. The sword concepts still needed revision before going to 3D, so the model is now the Studio starting choice with the earlier ones kept around.
The local images weren't good enough
In a matched chest comparison, the chest built from a fully local concept image didn't meet my bar for the game. Until a local source survives that test, the game uses a reviewed reference — sometimes from outside the local stack — with its origin recorded. The pipeline works, but a working pipeline isn't the same as a good-looking asset.
Sound, rigging, and a Studio to run it all
I added local sound-effect and music trials, and picked a generated hit and a combat loop to try in the game. A goblin auto-rig pilot produced a skinned model with test animations. To stop running everything from scripts, I built Asset Studio, a local app that runs each stage one job at a time and keeps a gallery of every run, including the failures.
Starting with one question
While working on game ideas I wondered how far I could get making the assets myself on local hardware, instead of hunting for them. The first version took an image to a 3D model using my Mac and a Windows laptop with an RTX 4080 doing the GPU work over SSH. A stylized boulder and a stone golem were the first models I imported into Unity.
