Scholarlink

<p>The rapid advancement of deep learning-based speech synthesis and voice cloning technologies has led to an increase in realistic deepfake voices, raising concerns about security, privacy, and the spread of misinformation. Detecting deepfake voices requires robust models that can distinguish between real and synthetic speech across various synthesis techniques. In this work, we propose a hybrid CNN+LSTM+GRU model that leverages Mel-Frequency Cepstral Coefficients (MFCC) and other speech features extraction for effective deepfake detection. The LSTM module learns long-term dependencies within the speech sequence, while the GRU enhances computational efficiency and mitigates vanishing gradient issues. The hybrid model achieves improved results in detecting synthetic speech generated by state-of-the-art text-to-speech and voice conversion systems, demonstrating superior performance over standalone CNN or RNN architectures. Experiments on the In-the-Wild dataset demonstrate a remarkable test accuracy of 99.84%, with precision, recall, and F1-scores all averaging 1.00. Cross-dataset evaluations on ASVspoof 5 and WaveFake confirm the robustness of the model, achieving more than 98% accuracy and an Equal Error Rate (EER) as low as 0.0013, with a throughput of 625.35 samples per second and a peak GPU memory usage of 2.01 GB (RTX 4060). This study highlights the effectiveness of hybrid deep learning architectures in combating voice spoofing and enhancing security in audio-based authentication systems. The proposed model aims to accurately distinguish between real and deepfake voices by utilizing the “In-the-Wild” dataset, which comprises both synthetic and authentic audio samples. The preprocessing stage extracts several features from the audio, including chroma_stft, rms, spectral centroid, spectral bandwidth, roll-off, and zero-crossing rate. In the proposed hybrid model, a CNN is employed to efficiently extract features from audio signals. At the same time, the temporal dependencies are sought out using GRU and LSTM layers, which process data sequentially.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hybrid Deep Learning Framework for Real and Deepfake Voice Detection

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Madhavanand Murty,
  • Shalini Tomar,
  • Shashidhar G. Koolagudi

摘要

The rapid advancement of deep learning-based speech synthesis and voice cloning technologies has led to an increase in realistic deepfake voices, raising concerns about security, privacy, and the spread of misinformation. Detecting deepfake voices requires robust models that can distinguish between real and synthetic speech across various synthesis techniques. In this work, we propose a hybrid CNN+LSTM+GRU model that leverages Mel-Frequency Cepstral Coefficients (MFCC) and other speech features extraction for effective deepfake detection. The LSTM module learns long-term dependencies within the speech sequence, while the GRU enhances computational efficiency and mitigates vanishing gradient issues. The hybrid model achieves improved results in detecting synthetic speech generated by state-of-the-art text-to-speech and voice conversion systems, demonstrating superior performance over standalone CNN or RNN architectures. Experiments on the In-the-Wild dataset demonstrate a remarkable test accuracy of 99.84%, with precision, recall, and F1-scores all averaging 1.00. Cross-dataset evaluations on ASVspoof 5 and WaveFake confirm the robustness of the model, achieving more than 98% accuracy and an Equal Error Rate (EER) as low as 0.0013, with a throughput of 625.35 samples per second and a peak GPU memory usage of 2.01 GB (RTX 4060). This study highlights the effectiveness of hybrid deep learning architectures in combating voice spoofing and enhancing security in audio-based authentication systems. The proposed model aims to accurately distinguish between real and deepfake voices by utilizing the “In-the-Wild” dataset, which comprises both synthetic and authentic audio samples. The preprocessing stage extracts several features from the audio, including chroma_stft, rms, spectral centroid, spectral bandwidth, roll-off, and zero-crossing rate. In the proposed hybrid model, a CNN is employed to efficiently extract features from audio signals. At the same time, the temporal dependencies are sought out using GRU and LSTM layers, which process data sequentially.

Graphical abstract