A Lightweight Log-Mel CNN for Audio Deepfake Detection: An Empirical Study of Speaker-Disjoint and Cross-Dataset Generalization

Abstract Synthetic speech generated by text-to-speech and voice-conversion systems can be difficult to distinguish from bonafide speech, creating a need for practical audio-spoofing detectors. This report describes a compact convolutional neural network (CNN) that classifies audio as real or fake using log-Mel spectrogram features. The supplied implementation combines the ASVspoof 2019 Logical Access training partition with the training partition of the In-the-Wild dataset. Audio is converted to mono, resampled to 16 kHz, cropped or zero-padded to four seconds, transformed using a short-time Fourier transform, projected onto 64 Mel bands, and normalized in the log-decibel domain. A four-block 2-D CNN is optimized for five epochs using binary cross-entropy with logits and AdamW. Model selection is performed using validation ROC-AUC, with the epoch-three checkpoint selected at 0.9890. On a reported speaker-disjoint In-the-Wild test set of 8,966 recordings, the checkpoint achieves ROC-AUC 0.99278, accuracy 98.32%, F1 98.79%, and an estimated equal error rate (EER) of 2.40%. However, an external WaveFake sample containing 700 real and 700 generated clips yields ROC-AUC 0.6630 and 50% accuracy at the unchanged 0.5 threshold; all clips are predicted as real at that threshold. This discrepancy highlights the difference between performance on a held-out split from a familiar dataset and transfer to a different synthesis domain. The system is therefore a promising baseline for the evaluated split, but the external result does not support claims of robust general-purpose audio deepfake detection.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-11729499/v1
Primary Topic
Speech Recognition and Synthesis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Lightweight Log-Mel CNN for Audio Deepfake Detection: An Empirical Study of Speaker-Disjoint and Cross-Dataset Generalization

Mayank Goswami, Gaurav Pratap Singh
1 links
Research Square
Speech Recognition and Synthesis
preprint

A Lightweight Log-Mel CNN for Audio Deepfake Detection: An Empirical Study of Speaker-Disjoint and Cross-Dataset Generalization

Mayank Goswami, Gaurav Pratap Singh
preprint en

Abstract

Abstract Synthetic speech generated by text-to-speech and voice-conversion systems can be difficult to distinguish from bonafide speech, creating a need for practical audio-spoofing detectors. This report describes a compact convolutional neural network (CNN) that classifies audio as real or fake using log-Mel spectrogram features. The supplied implementation combines the ASVspoof 2019 Logical Access training partition with the training partition of the In-the-Wild dataset. Audio is converted to mono, resampled to 16 kHz, cropped or zero-padded to four seconds, transformed using a short-time Fourier transform, projected onto 64 Mel bands, and normalized in the log-decibel domain. A four-block 2-D CNN is optimized for five epochs using binary cross-entropy with logits and AdamW. Model selection is performed using validation ROC-AUC, with the epoch-three checkpoint selected at 0.9890. On a reported speaker-disjoint In-the-Wild test set of 8,966 recordings, the checkpoint achieves ROC-AUC 0.99278, accuracy 98.32%, F1 98.79%, and an estimated equal error rate (EER) of 2.40%. However, an external WaveFake sample containing 700 real and 700 generated clips yields ROC-AUC 0.6630 and 50% accuracy at the unchanged 0.5 threshold; all clips are predicted as real at that threshold. This discrepancy highlights the difference between performance on a held-out split from a familiar dataset and transfer to a different synthesis domain. The system is therefore a promising baseline for the evaluated split, but the external result does not support claims of robust general-purpose audio deepfake detection.

Research Square
Indian Institute of Technology Roorkee (IN)
Speech Recognition and Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon