Introducing Open-Audiodeleto

Today we're open sourcing Open-Audiodeleto, an experimental audio-bypass risk classifier built by Clippsly.

Open-Audiodeleto analyzes audio files and estimates how closely they resemble known audio bypass uploads using acoustic features such as loudness patterns, spectral characteristics, silence ratios, and energy transitions. The project was created as an experiment in audio moderation research and machine learning.

During internal testing, Open-Audiodeleto achieved approximately 79% accuracy on a held-out test set, demonstrating that relatively simple machine learning techniques can identify meaningful differences between ordinary audio and known bypass-style uploads.

Unlike traditional audio moderation systems, Open-Audiodeleto does not identify songs, recognize artists, detect copyright ownership, or make moderation decisions. Instead, it produces a risk score that can be used as one signal among many during review.

Benchmark Results

To better understand how the model behaves across different types of audio, we tested Open-Audiodeleto against four categories:

• Confirmed Audio Bypasses
• Clippsly Releases
• AI Generated Music
• Public Domain Music

Average Risk Scores

The results show that confirmed bypass uploads consistently received the highest scores, while public-domain music and ordinary platform releases generally remained in the low-risk range.

Example Findings

• Confirmed Audio Bypasses scored significantly higher than Public Domain Music.
• Most Clippsly Releases remained in the low-risk range.
• AI Generated Music generally produced moderate-to-low risk scores.
• Public Domain Music consistently produced the lowest scores across the benchmark.

It's important to note that Open-Audiodeleto is not attempting to determine whether a song is copyrighted, identify artists, or detect ownership. The model is only estimating how closely an audio file resembles patterns observed in known bypass uploads.

Open Source Research

The project is fully open source and intended for researchers, developers, moderation teams, and anyone interested in understanding how audio classification systems can be built using traditional machine learning techniques.

Open-Audiodeleto should not be used as the sole basis for moderation actions. It does not prove copyright infringement, user intent, ownership, or policy violations.

This release is provided as a lightweight research model and reference implementation. The model is considered complete and is not expected to receive future updates.

We hope it serves as a useful learning resource for anyone exploring audio moderation, classification, and machine learning.

GitHub: https://github.com/Clippsly/open-audiodeleto