Authors:
Cesar Pachon and Dora Ballesteros.
Where published:
Big Data and Cognitive Computing, 2026, , Volume 10.
Abstract:
The paper introduces FakeVoiceFinder, a unified experimental framework for audio deepfake detection that systematically evaluates how different factors affect performance. Instead of focusing only on model architectures (like CNNs or Transformers), the authors design a platform that jointly analyzes model choices, hyperparameters (e.g., learning rate, epochs), and audio representations (e.g., spectrograms, scalograms, CQT). The framework allows researchers to run controlled and reproducible experiments, compare custom models with standard baselines under the same conditions, and study the interaction between model design and data representation. Overall, they provide a comprehensive benchmarking and evaluation tool that helps better understand what combinations of models and features work best for detecting synthetic audio.
Dataset names (used for):
- Fake Audio Dataset from the Mendeley dataset (ElevenLabs & Respeecher): used in the basic experiment.
- TTS/V2V Audio Deepfake Dataset: used for evaluating the inference module.
Some description of the approach:
The paper proposes FakeVoiceFinder, a unified framework for audio deepfake detection that evaluates how model architectures, training hyperparameters, and audio representations affect detection performance. It supports model-centric, data-centric, and hybrid benchmarking, enabling fair comparison between custom models and standard baselines under the same experimental conditions.
Keywords:
audio deepfake; synthetic audio; CNN; Transformers; model-centric; data-centric; TTS; V2V.
Instance Represent:
Each instance is an audio sample labeled as natural/real or AI-generated/fake. In the framework, audio samples are transformed into time-frequency tensors such as mel-spectrograms, log-spectrograms, scalograms, or CQT representations before being passed to a model.
Dataset Characteristics:
The main example dataset is balanced with 600 synthetic and 600 real audios. The synthetic source includes 492 V2V and 108 TTS samples generated by ElevenLabs and Respeecher. It includes male and female voices, audio duration of 8–10 seconds, and sampling rate of 22,050 Hz.
Subject Area:
Synthetic audio detection, audio deepfake detection, audio forensics, machine learning benchmarking, CNN/Transformer-based detection, data-centric and model-centric evaluation.
Associated Tools:
FakeVoiceFinder framework; Python modules such as prepare_dataset.py, model_loader.py, trainer.py, experiment.py, metrics.py, and inference.py; architectures including AlexNet, ResNet, VGG, DenseNet, MobileNet, EfficientNet, SqueezeNet, Inception, ConvNext, ViT.
Feature Type:
The framework supports four time-frequency feature representations: mel-spectrogram, log-spectrogram, scalogram / DWT, and CQT.
Number of Instances:
For the basic experiment: 1,200 total audio samples, including 600 synthetic and 600 real. The synthetic Mendeley dataset itself contains 600 synthetic audios: 492 V2V and 108 TTS. The inference evaluation separately uses 120 samples: 60 natural and 60 fake.
Number of Features:
The paper does not report a single fixed number of input features, because the feature dimensionality varies depending on which transformation and hyperparameters the user configures. All representations are then resized to a user-defined cfg.image_size (e.g., 224×224) before being fed into the models.