‘Local AI’ is a collection of AI tools that run entirely on your computer.

There is no cloud account, no subscription, no usage fees, and no per-prompt limits.

Your prompts and results stay on your machine.

It consists of the following assets:

• 🎵 Music, compose songs offline from a style preset and optional lyrics.
• 🗣️ Voice, generate speech offline from a Voice Preset, with optional emotion tags.
• 🛡️ Moderator, moderates chats in more than 23 languages.
• and more to come…

Requirements  

All ‘Local AI’ assets are developed for Unity 6 (version 6000.0 or higher).

Music and Voice run in the Windows x86_64 Editor. They are not added to a player build unless you import the generated audio yourself.

Moderator is a runtime component: it runs in play mode and in player builds (including WebGL).

A GPU is recommended. It is much faster than CPU.

If you don’t know how to install Unity 6, follow this official tutorial.

Installation  

1. Import the assets into your Unity project. If you’ve purchased the bundle, you’ll be able to download all the assets from the store at no charge.
2. Open editor tools from Window → Fronkon Games → Local AI, or add a runtime component such as Moderator.
3. Download extra model files when the asset asks for them (Music, Voice). Moderator includes its Sentis model in the package.

🎵 Music  

‘Local AI: Music’ generates songs offline inside the Unity Editor. Describe style and arrangement, optionally write sectioned lyrics, then render a WAV on your machine.

• Private and local. Prompts, lyrics, and the WAV never leave your computer.
• Editor-only. Tools live under Window → Fronkon Games → Local AI → Music. Nothing is added to a player build unless you import the generated audio yourself.
• Hundreds of styles. Genre presets ship in the package, including instrumental variants. Duplicate and edit them, or create your own.
• Optional lyrics. Assign a lyric asset for a sung track. Leave lyrics empty (or use a style that describes no singer) and generation is treated as instrumental without rewriting the style asset.
• WAV output. Stereo, 44.1 kHz, 16-bit PCM. Length 5–300 s.
• Game-ready loops. Enable Seamless loop to search for a natural wrap in the WAV and mark the imported AudioClip as loopable.
• CUDA, then Vulkan, then CPU. A GPU is not required. CUDA is used when a compatible NVIDIA GPU is present.

Requisites  

To ensure optimal performance and compatibility, your project must meet the following requirements:

• Unity: 6000.0 or higher.
• Editor: Windows x86_64.
• GPU: not required. CUDA is used when a compatible NVIDIA GPU is present, then Vulkan, then CPU.
• VRAM: about 12 GB for the included quantized models. A 16 GB card is more comfortable.
• Free space: at least 10 GB of free space.

Installation Guide  

1. Import Local AI: Music.
2. Create a music style using the Window → Fronkon Games → Local AI → Music → Style tool (or use or modify one of the hundreds of included presets).
3. Create the lyrics using the Window → Fronkon Games → Local AI → Music → Lyrics tool.
4. Open Window → Fronkon Games → Local AI → Music → Generator. Select the style and lyrics.
5. Set the Models folder and click Download, or select a folder that already contains the five GGUF files.
6. Click Generate.

Tools  

Style  

A Style is a MusicPreset asset: a structured description of genre, mood, vocals, and arrangement. The generator turns this text into the caption sent to the model.

Create one with New, pick an existing asset in the object field, or start from a preset under Editor/Presets/. Changes are saved on the asset when you edit it.

Global  

Vocal Details  

These fields describe a singer. If you generate without lyrics (or with empty section text), the generator builds an instrumental request and does not write these fields back onto the asset.

Arrangement  

Lyrics  

A Lyrics asset is a MusicLyric: a title plus reorderable sections. Lyrics are optional. If you skip this asset, or every section has empty text, the generator treats the song as instrumental.

Multiple voices (experimental)  

To assign different singers to different sections (for example, a duet) describe both voices in Style → Vocal Details (gender, timbre, and who sings which part).

Add short voice tags on their own lines inside the section text: [male vocal], [female vocal], or [duet].

Caption-only instructions are usually not enough, tags in the lyrics matter too. Keep each tag to one or two words, longer tags may be sung aloud.

[Verse]
[male vocal]
I walked alone through midnight rain
Every street remembers your name

[Verse]
[female vocal]
You left a letter by the door
Said you weren't coming back no more

[Chorus]
[duet]
We were fire, we were flame
Nothing left but ash and shame

In the Style, name both singers and state who opens, for example, a deep male baritone on the first verse, a bright female soprano on the second, both joining the chorus, with no choir or doubled backing vocals.

Section-level switching works best; line-by-line alternation usually fails. Casting is not deterministic, try different seeds and check the first few seconds; if the wrong voice opens, generate again.

Generator  

Assign a style, optionally lyrics, then Generate. Native backends are tried in order: CUDA → Vulkan → CPU.

Diffusion  

Language Model  

Generation  

Powered by MiniMax Music 3.

🗣️ Voice  

‘Local AI: Voice’ generates speech offline inside the Unity Editor. Invent a speaker (or clone a clip), bake it onto a Voice Preset, then speak lines in Voice Generator. Optional emotion tags change delivery without changing the speaker.

• Private and local. Prompts, clips, and the WAV never leave your computer.
• Editor-only. Tools live under Window → Fronkon Games → Local AI → Voice. Nothing is added to a player build unless you import the generated audio yourself.
• Voice Preset. Instruction invents a speaker from gender and a description. Clone copies a reference clip and its transcript. Generate bakes a dry 24 kHz take onto the asset.
• Voice Generator. Always clones that baked speaker. Tag emotion per block.
• WAV output. Mono, 24 kHz, 16-bit PCM.
• Only English supported. Support for Chinese is experimental.
• Post-FX. Optional mix, space, polish, and stylized effects live on the preset. Play in Voice Preset and Generate in Voice Generator apply them. The bake stays dry so the clone reference is unchanged.
• CUDA, then Vulkan, then CPU. A GPU is not required. CUDA is used when a compatible NVIDIA GPU is present.

Requisites  

To ensure optimal performance and compatibility, your project must meet the following requirements:

• Unity: 6000.0 or higher.
• Editor: Windows x86_64.
• GPU: not required. CUDA is used when a compatible NVIDIA GPU is present, then Vulkan, then CPU. A GPU is strongly recommended.
• VRAM: about 8 GB. A 12 GB card is the minimum recommended configuration.
• Disk: about 8 GB for model files. At least 10 GB of free space.

Installation Guide  

1. Import Local AI: Voice.
2. Open Window → Fronkon Games → Local AI → Voice → Voice Preset.
3. Set the Models folder and click Download, or select a folder that already contains Voice.gguf.
4. Create or pick a Voice Preset. Fill Instruction (gender + description) or Clone (clip + transcript).
5. Click Bake voice preset.
6. Open Window → Fronkon Games → Local AI → Voice → Voice Generator. Assign the preset, write the line, click Generate voice.

Tools  

Voice Preset  

A Voice Preset is a VoicePreset asset: the speaker definition plus baked 24 kHz samples. Create one with New, or pick an existing asset in the object field (+150). Changes are saved on the asset when you edit it.

Generate speaks a fixed test line and stores the dry samples on the preset. Play previews those samples with the preset’s post-FX. The stored bake stays dry for cloning.

Voice  

Sampling  

How the bake samples the speaker. Stable take (temperature 0, depth temp 0, top-k 1) repeats more closely with a fixed seed. Default uses CFG for this mode, temperature 0.9, depth temp 0.9, top-k 50, top-p 1, and invents more variation.

Bake voice  

Audio effects  

Optional post-FX stored on the Voice Preset. Voice Generator applies them after creation. Play in Voice Preset applies them to the preview. The bake stays dry.

Mix / loudness  

Space / tone  

Master polish  

Stylized  

Voice Generator  

Assign a baked Voice Preset, write the spoken text, then Generate voice. Emotion tags are applied per block. Unknown names and unclosed tags are errors. Tags are not nested. These are the supported tags:

Insert a pause with <silence>1</silence> (greater than 0, up to 5 seconds). Example:

Welcome to the keep.
<calm>The next checkpoint is quiet.</calm>
<silence>1</silence>
<happy>The reward is yours.</happy>

Powered by a LORA of Vox CPM 2.

🛡️ Moderator  

‘Local AI: Moderator’ classifies chat text on-device in runtime. Drop a Moderator component on a GameObject, then call Evaluate, IsFlagged, or Filter to score hateful, abusive, and threat content plus the protected group it appears to target.

• Private and local. Messages never leave the player’s machine. No cloud account and no API keys.
• Runtime. This asset is meant for play mode and player builds.
• ~15 ms per message on desktop GPUCompute. CPU is supported. WebGL falls back to GPUPixel, then CPU, and is much slower (several seconds).
• 23 languages: English, German, French, Dutch, Portuguese, Spanish, Italian, Polish, Danish, Swedish, Finnish, Hungarian, Czech, Slovak, Slovenian, Croatian, Bulgarian, Greek, Romanian, Lithuanian, Latvian, Estonian, Irish.
• HateCheck F1 0.87 (recall 0.96) on 29,797 cases. English HateCheck F1 0.86.
• Three content scores: hateful, abusive, threat. Ten target groups: race, colour, religion, descent, national/ethnic origin, sexual orientation, gender, disability, age.
• About 160 MB on disk for the included quantized .sentis model and tokenizer.json.

Requisites  

To ensure optimal performance and compatibility, your project must meet the following requirements:

• Unity: 6000.0 or higher.
• Unity Inference Engine 2.6.1 (com.unity.ai.inference). The package pulls this in as a dependency.
• GPU: recommended. Default backend is GPUCompute. CPU works. WebGL cannot run GPUCompute (no compute shaders). It tries GPUPixel, then CPU.
• Disk: about 160 MB for the model and tokenizer (includes a more accurate, non-quantized model).

Installation Guide  

1. Import Local AI: Moderator.
2. Add Fronkon Games → Local AI → Moderator to a GameObject. Empty model and tokenizer fields are filled from the package when possible.
3. Open the demo scene at Assets/FronkonGames/LocalAI/Moderator/Demo/Moderator_demo.unity, or try the WebGL demo.

Default thresholds are 0.40 (hateful / target) and 0.50 (abusive / threat). Input is truncated at 128 tokens.

Component  

Model  

Thresholds  

A category is flagged when its score is at or above the slider. Flagged is true when any content score (hateful, abusive, or threat) is over its limit.

Debug  

The documentation link at the bottom of the inspector opens this page.

API  

Namespace: FronkonGames.LocalAI.Moderator.

using FronkonGames.LocalAI.Moderator;

Moderator moderator = GetComponent<Moderator>();

if (moderator.IsReady == false)
  moderator.Load();

if (moderator.IsFlagged(message) == true)
{
  // Drop, mute, or replace the line.
}

ModeratorResult result = moderator.Evaluate(message);
if (result.Flagged == true)
{
  float hate = result.GetScore(ModeratorContent.Hateful);
  ModeratorTarget target = result.TopTarget;
}

string shown = moderator.Filter(message, "[blocked]");

Non-blocking (yields one frame so UI can show a busy state):

ModeratorResult result = await moderator.EvaluateAsync(message);

IsBusy is true while an evaluation is running. A second call at the same time throws.

How the demo flags messages  

Moderator.Evaluate returns scores. result.Flagged is only one policy: any content head (hateful, abusive, threat) at or above its threshold.

The included demo does not stop there. It applies a small extra rule on top of those scores (if a content score is close to its limit, or a target group is very high, it boosts the strongest content score using how many target heads also fire). That is demo-only. It is not used by IsFlagged, Filter, or result.Flagged on the component.

There is no perfect way to turn scores into a yes/no. Chat, lobbies, kids’ games, and competitive servers all need different trade-offs (catch more hate vs. fewer false positives). Use the scores, thresholds, and target heads to write your own algorithm for that environment. The demo is an example, not a recommendation for shipping.

Languages  

Trained for 23 languages. The demo and HateCheck eval cover English, German, French, Dutch, Portuguese, Spanish, Italian, and Polish. Other listed languages can work but are not scored in that suite.

WebGL  

WebGL cannot create a GPUCompute worker. The component tries GPUPixel, then CPU. Expect several seconds per message in the browser. On desktop GPUCompute it is about 15 ms.

F.A.Q.  

Does Moderator need an internet connection?  

No. Inference runs on the device with Unity Inference Engine. Nothing is sent to a server.

Why is the WebGL demo slow?  

WebGL has no compute shaders, so GPUCompute is not used. The demo falls back to GPUPixel or CPU. Expect several seconds per message in the browser. On a desktop GPU it is about 150 ms.

Will I run into any copyright issues?  

Generated files are yours to ship in games or media (commercial or not).

You shouldn’t use this assets for other purposes besides generating assets for games or promotional material.

You cannot impersonate real people, plagiarize copyrighted works, or use it for illegal purposes.

It’s your responsibility.

Do I need to say that my game use AI?  

It depends on the platform.

For example, on Steam you must disclose the use of generative AI if any AI-generated content or assets actually ship with your game or appear in your store and marketing materials.

Support  

Do you have any problem or any suggestions? Send me an email to fronkongames@gmail.com and I’ll be happy to help you.

If you want to report an error, it helps a lot if you include the Unity log file.