secureaudio-dual-expert-1d

rijal028/secureaudio-dual-expert-1d

secureaudio-dual-expert-1d is a machine learning model from rijal028. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as apache-2.0.

A physics-driven 1D Multi-Scale CNN (~8.4 MB, ~2,100,000 parameters) trained 100% from scratch on raw 1D acoustic waveforms (8,000 samples / 500 ms @ 16 kHz).

Downloads · 30 days

Access

Public

Updated Oct 6, 2026

Repo size

37.4 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.pt27.2 MB · 100%

At a glance

Library: pytorch
License: apache-2.0
Access: Public
Created: Sep 23, 2026
Updated: Oct 6, 2026
SHA: 60ec092d

Model card

From the Hugging Face model README

SecureAudio 2.0 Pro: High-Capacity 1D Acoustic Forensics (~2.1M Params)

A physics-driven 1D Multi-Scale CNN (~8.4 MB, ~2,100,000 parameters) trained 100% from scratch on raw 1D acoustic waveforms (8,000 samples / 500 ms @ 16 kHz).

Equipped with a triple-branch front-end modeling macro-vocal dynamics, mid-formant transitions, and micro-transient digital vocoder phase anomalies under forced social media compression (WhatsApp Opus 16–32 kbps).

1. Large-Scale Blind Cross-Dataset Benchmark (N = 1,000 Unseen External Audio)

Tested on unseen voice generators (ElevenLabs, Hume AI, Kokoro TTS, Speechify, Amazon Polly) against natural human speech, evaluated under forced WhatsApp Opus 24 kbps degradation:

Metric	Score	Sample Ratio
Verified Accuracy	97.50%	975 / 1,000
False Alarm (Real Speech -> AI)	1.20%	6 / 500
Missed AI (Synthetic -> Real)	0.40%	2 / 500
Quarantined Fail-Safe (Inconclusive)	1.70%	17 / 1,000

2. In-Corpus Holdout Benchmark (N = 374 Files, Forced WA Opus 24k)

Verified Accuracy: 97.06% (363 / 374 files)
False Alarm: 0.80% (3 / 187 files)
Missed Deepfakes: 0.00% (0 / 187 files)
Inconclusive: 2.14% (8 / 374 files)

Library: pytorch
License: apache-2.0
Created: Sep 23, 2026
Updated: Oct 6, 2026