Updates

The v5 multi-source training approach resolved the critical external false positive problem identified in v2. By training on diverse benchmark voice corpora beyond the original baseline, the model now correctly identifies real human speech from previously unseen recording environments.

Problem

Voice SSL v2 achieved strong detection metrics (EER 1.0%, 99.54% detection rate) but exhibited a high false positive rate on external bonafide datasets — real human recordings from unseen environments were frequently misclassified as AI-generated. Root cause: the model had only seen a narrow recording environment during training.

Solution

v5 training incorporated bonafide samples from multiple benchmark voice corpora, totaling tens of thousands of samples with domain-weighted sampling to maintain source balance.

Results

Training reproducibility was verified across 4 independent PCs with different hardware configurations. The pc4 model (Ultra 9 285K + RTX 5060) was selected as the v5 reference model.

Large-scale evaluation of the Voice SSL v2 engine across 5,328 TTS samples confirmed robust detection under real-world telephony conditions. The test covered 4 commercial TTS engines, 32 voices, and 8 phone environment simulations.

Test Matrix

• •TTS Engines: ElevenLabs, OpenAI TTS-1, Google Cloud Neural2, Microsoft Edge TTS
• •Languages: English, Korean, Japanese, Chinese
• •Phone Conditions: G.711 codec, 8kHz resampling, background noise (SNR 10dB), short utterances (2 seconds), combined phone environment

Key Findings

A counterintuitive finding: phone codec artifacts (G.711, 8kHz resampling) actually amplify AI signal evidence, making detection easier under telephony conditions than in clean audio. All 115 missed detections were English-language clean audio samples.

USPTO provisional patent specification completed, covering three core inventions: the multi-stage Audio verification engine, the Voice deepfake detection engine, and the unified detection architecture that combines both for music containing vocal content.

Coverage

• •Audio Engine: Multi-stage verification with independent analysis engines and cross-validation consensus
• •Voice Engine: Self-supervised learning approach for deepfake voice detection optimized for telephony conditions
• •AI Music + Voice Unified Detection: Combined architecture applying both audio and voice analysis to vocal-containing music for comprehensive verification

The Voice deepfake detection engine reached production readiness, achieving EER 1.0% on a standard benchmark voice evaluation set.

Architecture Decision

The original voice PoC achieved functional but suboptimal results (EER 2–6%). Analysis of published research informed a major architecture upgrade that improved performance by 10–50x on equal error rate.

Performance

All 13 unseen attack systems in the benchmark voice evaluation set were detected at 97.5% or above. An alternative graph attention backend was evaluated and rejected — it exhibited overfitting and worse performance than the simpler Attention Pooling + MLP approach.

Real-World TTS Validation

Initial validation against 55 commercial TTS samples (ElevenLabs, OpenAI, Google Cloud — 13 voices, 2 languages) achieved 100% detection rate. The model was trained only on academic benchmark voice corpora but successfully generalized to 2025–2026 commercial TTS engines.

DetectX Audio now operates exclusively in Enhanced Mode, a dual-engine verification architecture designed to maximize human protection while maintaining effective AI detection.

Architecture Overview

Enhanced Mode combines two complementary engines working in sequence:

• •DetectX Engine (Primary): A deep learning engine trained on millions of verified human music samples. Optimized for near-zero false positives. If the DetectX Engine determines content is human, the verdict is trusted immediately.
• •Verification Engine (Secondary): Activates when the DetectX Engine score exceeds the threshold. Performs multi-layer analysis to boost AI detection accuracy.

Performance Characteristics

• •Human False Positive Rate: <1% — human creators are protected
• •AI Detection Rate: Strong detection for confirmed AI-generated content
• •Binary Verdicts: No probabilistic scores, only structural observations

Design Philosophy

The dual-engine approach prioritizes human safety as a hard constraint. By using the DetectX Engine as the primary filter, the system ensures that human creative work is never unfairly flagged. The DetectX Deep Forensic Engine serves as a secondary check only when the primary engine indicates potential AI content.

After extensive testing across multiple baseline construction approaches, the minimal strategy has been locked for production deployment.

Decision Summary

The human baseline will be constructed using a minimal, high-confidence corpus rather than an expansive, diverse corpus. This decision prioritizes false positive prevention over detection sensitivity.

Rationale

• •Larger baselines increase the risk of including edge-case human content that resembles AI patterns, leading to baseline contamination.
• •Minimal baselines with strict provenance verification provide cleaner separation between human and AI signal geometry.
• •False positives (human work flagged as AI) cause more harm than false negatives (AI work not detected). The minimal strategy optimizes for human safety.

Implementation

The DetectX Engine has been trained on millions of verified human-created audio samples spanning diverse genre categories. Each sample has documented provenance including recording session metadata, artist verification, and production chain attestation.

Validation Results

Testing against a held-out validation set of 800 verified human samples showed zero false positives. Testing against a corpus of 1,200 AI-generated samples showed 94.2% detection rate.

The 5.8% of AI samples not detected exhibited signal geometry within human baseline parameters. These samples are being analyzed to determine whether baseline expansion is warranted or whether they represent legitimate edge cases.

Next Steps

• •Deploy minimal baseline to production verification pipeline
• •Monitor false positive reports and baseline performance metrics
• •Continue analysis of undetected AI samples for potential baseline refinement
• •Document baseline versioning and update procedures