DecodeX Base: Real-Time Neural Audio Engine
This is the primary foundational model for the DecodeX audio verification platform, running in parallel with V2 to provide highly accurate, real-time deepfake and synthetic voice detection. It achieves the following results on the evaluation set:
- Loss: 0.0829
- Accuracy: 0.9882
Model description
DecodeX Base operates as the core Wav2Vec 2.0 layer in our dual-engine cross-verification architecture. It is optimized for high-speed inference on raw PCM audio streams, specifically designed to identify zero-day synthetic voices and voice cloning attempts before they breach communication networks.
Intended uses & limitations
Intended for deployment in enterprise PBX systems, SIP trunks, and local edge hardware to act as a first line of defense against AI-generated social engineering and spoofing attacks.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 3e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 5
Training results
| Training Loss | Epoch | Step | Accuracy | Validation Loss |
|---|---|---|---|---|
| 0.1448 | 1.0 | 1900 | 0.9601 | 0.1447 |
| 0.0673 | 2.0 | 3800 | 0.9824 | 0.0817 |
| 0.0178 | 3.0 | 5700 | 0.9796 | 0.1054 |
| 0.0002 | 4.0 | 7600 | 0.9824 | 0.1074 |
| 0.0108 | 5.0 | 9500 | 0.9882 | 0.0829 |
Framework versions
- Transformers 4.39.3
- Pytorch 2.1.2
- Datasets 2.18.0
- Tokenizers 0.15.2