Totally Human is a native desktop app for detecting AI-generated music. It analyzes audio files for spectral artifacts and statistical anomalies that are typical of synthetically produced music, and returns the result as a percentage estimate: how much of the analyzed music is likely AI-generated versus human-made.
The app is bilingual (German / English) and the language can be switched at any time in the settings — the interface changes immediately (live) and the selection is saved for the next launch.
This repository contains the source code for both platforms. The compiled, ready-to-run builds of the app are published as GitHub Releases (see the Releases page).
| Platform | Source code | Compiled build |
|---|---|---|
| macOS | macOS Source/ |
Totally Human.app (in Releases) |
| Windows | Windows Source/ |
Totally Human.exe (in Releases) |
The
Totally Human.appandTotally Human.exefiles are already compiled, runnable versions of the app — no build required. Download them from the Releases page.
- File Analysis — Single-file analysis via drag & drop or file selection. Shows AI probability, artifact score, confidence, classification, and additional information.
- Batch Analysis — Multiple files at once; tabular results and a summary (AI / human).
- Training — Add your own examples (real and AI-generated music), retrain the model, and import seed data. All data is stored locally.
- Visualization — Spectrum, baseline, and fingerprint displayed as charts, along with segment strengths.
- Settings — Language, theme, quality threshold, segment length, storage folder, and seed re-import.
The app detects AI-generated music using several methods:
- Fourier artifact analysis — regular peaks between 5–16 kHz
- Self-similarity matrix (MFCC-based)
- Splice detection — temporal inconsistencies
- Obfuscation detection
The result is displayed as a percentage, estimating how much of the analyzed music is AI-generated versus human-made.
Recommended thresholds: AI ≥ 0.6 · human ≤ 0.4 · otherwise manual review.
The app ships with a pre-trained base model (seed data: 148 labeled examples) so it works out of the box. This base is only a starting point — it is not a finished, all-knowing detector. The real strength of Totally Human is that you can train it yourself and continuously improve it.
- Add your own examples — feed the app real (human-made) music and complete AI-generated music as labeled training examples.
- Retrain the model — after adding examples, retrain the model so it learns the acoustic patterns of your music collection and the AI generators you care about.
- Import seed data — re-import the bundled seed data at any time to restore or extend the base.
- Incremental learning — the training database is expanded with every training session. The more diverse and representative the examples you add, the more accurate the detector becomes over time.
No single pre-trained model can know every AI generator, every genre, and every production style. AI music tools evolve constantly, and new generators appear regularly. By training the app on your own data, you:
- Adapt it to your use case — your genres, your sources, your quality standards.
- Keep it current — teach it to recognize new AI generators as they appear.
- Improve accuracy — the model gets better with every training session, because it learns from real, labeled examples rather than relying on a fixed base.
For the best results, use real, human-made music and complete, unmodified AI-generated tracks as your labeled examples:
- Real music — tracks you know are human-made, ideally covering the genres you want to analyze.
- AI-generated music — complete AI-generated tracks (not fragments), ideally from the generators you want to detect.
The clearer and more consistent your labels are, the more reliable the trained model becomes.
All training happens locally on your device. Training data, the trained model, and settings are stored only in your local user data directory (see below). The app does not upload, share, or collect your training data or your audio files.
You are responsible for the audio you use for training. Please only use content you are legally entitled to use:
- Your own music — tracks you created or own.
- Properly licensed content — music you have a license or permission to use for this purpose.
- Freely available content — music released under a license that permits use (e.g. Creative Commons, public domain).
Training is a local, private process: the app analyzes the audio on your machine and stores only the derived model locally. It does not redistribute, publish, or share the audio or the trained model. However, using copyrighted music without permission — even for private analysis — may still be restricted by the rights holder, so always make sure you have the necessary rights before adding content to the training set.
- macOS 13.0 (Ventura) or later
- Xcode 15+ (only required to build from source)
- Windows 10/11 (64-bit)
- No separate .NET installation required — the EXE is self-contained (the .NET runtime is embedded) and runs with a double-click.
Training data, the trained model, and settings are stored locally:
- macOS:
~/Library/Application Support/TotallyHuman/ - Windows:
%APPDATA%\TotallyHuman\
On first launch, embedded seed data (model_seed.json, training_seed.json, 148 examples) is automatically imported. The training database is expanded incrementally with each training session — this is the base model described in Training & Continuous Learning.
- Open
TotallyHuman.xcodeprojin Xcode - Signing: select your Team or use "Sign to Run Locally"
- Press
Cmd+Rto run the app
Option A — Automatically via GitHub Actions (recommended): The repository contains the workflow .github/workflows/build.yml. On every push to main/master (or manually via Actions → Build Windows EXE → Run workflow), a windows-latest runner builds the portable EXE and makes it available as the artifact TotallyHuman-Windows.
Option B — Locally on Windows: Requires the .NET 8 SDK.
dotnet publish TotallyHuman/TotallyHuman.csproj -c Release -r win-x64 --self-contained true \
-p:PublishSingleFile=true -p:EnableCompressionInSingleFile=true -o publishResult: publish\TotallyHuman.exe (portable, no installation required).
Note: WPF is Windows-specific.
dotnet build/publishonly works on Windows (or on the GitHub Actionswindows-latestrunner). On Linux/macOS the C# code can be checked, but a working EXE cannot be produced.
- macOS: Swift / SwiftUI, Apple vDSP (FFT)
- Windows: .NET 8 / WPF (
net8.0-windows), MVVM-style architecture withAppStateas observable state- NAudio + NAudio.Vorbis for audio decoding (MP3, WAV, AIFF, M4A, FLAC, OGG, OPUS, WMA, ALAC → Mono, 44.1 kHz, normalized)
- OxyPlot.Wpf for charts
- Manual radix-2 FFT (replacing Apple's vDSP), ensemble using FFT sizes 2048/4096/8192
If you like the project and want to support its development:
- Support the project: https://ko-fi.com/totallyhumanapp/
- GitHub: https://github.com/TotallyHumanApp/
- Instagram: https://www.instagram.com/totallyhumanapp/
- Contact: totallyhumanapp@gmail.com
All rights remain with the author. This repository is provided without a license — the use, reproduction, and redistribution of the code is not permitted without explicit permission.