Latin-American voice anti-spoofing dataset

Abstract

This dataset contains samples of spoof and real human voice with different accents from Latin-American countries. Table 1. Real samples distribution <strong>Accent</strong> <strong>Gender</strong> <strong># Speakers</strong> <strong># Files</strong> <strong>Nomenclature</strong> Colombian Male Female 17 14 2534 2070 com cof Chilean Male Female 17 12 2487 1602 clm clf Peruvian Male Female 20 18 2917 2529 pem pef Venezuelan Male Female 12 10 1754 1463 vem vef Argentinian Male Female 12 30 1670 3790 arm arf Total 162 22816 The bonafide samples were obtained from the following sources: Colombian accents: https://www.openslr.org/72/ (License) Chilean accents: https://www.openslr.org/71/ (License) Peruvian accents: https://www.openslr.org/73/ (License) Venezuelan accents: https://www.openslr.org/75/ (License) Argentinian accents: https://www.openslr.org/61/ (License) The strategies used to generate the spoof samples: Table 2. Spoof Samples distribution <strong>Name</strong> <strong>Type</strong> <strong>#Samples</strong> StarGAN Voice conversion 16000 CycleGAN Voice conversion 16000 Diffusion Voice conversion 16000 TTS Text-to-speech 5000 TTS-StarGAN Text-to-speech / Voice conversion 2500 TTS-Diff Text-to-speech / Voice conversion 2500 StarGAN-VC: Non-parallel many-to-many Voice Conversion Using Star Generative Adversarial Networks Cyclegan-VC: Non-parallel voice conversion using cycle-consistent adversarial networks Diffusion-Based Voice Conversion with Fast Maximum Likelihood Sampling Scheme TTS: Microsoft azure TTS TTS-VC: Microsoft azure TTS + StarGAN/Diff Table 3. Dataset overview <strong>Audio Samples</strong> <strong>Human Speakers</strong> <strong>Spoofing algorithms</strong> <strong>Sampling rate</strong> Bonafide Spoof 22816 58000 Male Female 78 84 VC TTS VC and TTS 3 1 2 16kHz On the <em>protocol.txt</em> file is listed all the files with the following structure: <em> Subject_id file_name – spoof_type Label</em> Consider this line on protocol.txt file:<br> <em>arf_00295 StarGAN-arf _00295_01349969200-cof _03349 _0077577 - StarGAN spoof</em><br> The first part (arf_00295) represent the subject id, from which we can also identify the accent and the gender (see nomenclature column on Table 1). The file name identify the type of spoof following for the source audio file and the target file. StarGAN represents the type of spoof. According to the table 2, this method is a Voice Conversion algorithm. If the file is a bonafide sample, we replace the <em>spoof_type</em> with a dash (-). Finally at the end of the line we refer the kind of label of the file, in the example, the file corresponds to a spoof case. Each zip file contains 6 folders, each one holds a type of samples. For the voice conversion folders, there are 25 sub-folders that indicate the conversion between accents. For example, Argentina-Venezuela folder indicates that the source accent of the file is Argentinian and the target is Venezuelan accent. Inside the folder there are 64 sub-folders that represent the subjects used for the conversion. For instance, the folder arf_00295-vem_04310 means that the source is an Argentinean female and the target is a Venezuelan male (see Table 1 for nomenclature). In the case of a Text-to-Speech folder there are 5 sub-folders that represent the accents. A TTS-VC folder there are 2 sub-folders that represent the voice conversion strategy used. Inside there are other sub-folders for the different combinations of source and target accents. You can check the folder tree structure in the tree.txt file. Table 3 shows a summary of the resulting dataset.