Codecfake dataset - training set (part 1 of 3)

<p>This dataset is the training set (part 1 of 3) of the Codecfake dataset , corresponding to the manuscript "<a href="https://arxiv.org/abs/2405.04880"><strong>The Codecfake Dataset and Countermeasures for Universal Deepfake Audio Detection</strong></a>".</p> <h1>Abstract</h1> <p>With the proliferation of Audio Language Model (ALM) based deepfake audio, there is an urgent need for effective detection methods. Unlike traditional deepfake audio generation, which often involves multi-step processes culminating in vocoder usage, ALM directly utilizes neural codec methods to decode discrete codes into audio. Moreover, driven by large-scale data, ALMs exhibit remarkable robustness and versatility, posing a significant challenge to current audio deepfake detection (ADD)<br>models. To effectively detect ALM-based deepfake audio, we focus on the mechanism of the ALM-based audio generation method, the conversion from neural codec to waveform. We initially construct the Codecfake dataset, an open-source large-scale dataset, including two languages, millions of audio samples, and various test conditions, tailored for ALM-based audio detection. Additionally, to achieve universal detection of deepfake audio&nbsp;and tackle domain ascent bias issue of original SAM, we propose<br>the CSAM strategy to learn a domain balanced and generalized&nbsp;minima. Experiment results demonstrate that co-training on&nbsp;Codecfake dataset and vocoded dataset with CSAM strategy yield&nbsp;the lowest average Equal Error Rate (EER) of 0.616% across all&nbsp;test conditions compared to baseline models.</p> <h1>Codecfake Dataset</h1> <p>Due to platform restrictions on the size of zenodo repositories, we have divided the Codecfake dataset into various subsets as shown in the table below:</p> <table> <tbody> <tr> <td>Codecfake dataset</td> <td>description</td> <td>link</td> </tr> <tr> <td>training set (part 1 of 3) &amp; label</td> <td>train_split.zip &amp; train_split.z01 - train_split.z05</td> <td><a href="https://zenodo.org/records/13838106"><strong>https://zenodo.org/records/13838106</strong></a></td> </tr> <tr> <td>training set (part 2 of 3)</td> <td>train_split.z06 - train_split.z10</td> <td><a href="https://zenodo.org/records/13841652">https://zenodo.org/records/13841652</a></td> </tr> <tr> <td>training set (part 3 of 3)</td> <td>train_split.z11 - train_split.z16</td> <td><a href="https://zenodo.org/records/13853860">https://zenodo.org/records/13853860</a></td> </tr> <tr> <td>development set</td> <td>dev_split.zip &amp; dev_split.z01 - dev_split.z02</td> <td><a href="https://zenodo.org/records/13841216">https://zenodo.org/records/13841216</a></td> </tr> <tr> <td>test set (part 1 of 2)</td> <td>Codec test: C1.zip - C6.cip &amp; ALM test: A1.zip - A3.zip</td> <td><a href="https://zenodo.org/records/13838823">https://zenodo.org/records/13838823</a></td> </tr> <tr> <td>test set (part 2 of 2)</td> <td>Codec unseen test: C7.zip</td> <td><a href="https://zenodo.org/records/11125029">https://zenodo.org/records/11125029</a></td> </tr> </tbody> </table> <h1>Countermeasure</h1> <p>The source code of the countermeasure and pre-trained model are available on GitHub <a href="https://github.com/xieyuankun/Codecfake">https://github.com/xieyuankun/Codecfake</a>.<br><br>The Codecfake dataset and pre-trained model are licensed with CC BY-NC-ND 4.0 license.</p>