ReplayFake-CN

ReplayFake-CN is a large-scale, multi-device Chinese dataset for physically replayed deepfake detection, comprising 184,954 utterances (232.5 hours) replayed synthetic utterances. It contains speech generated by 22 TTS models, with texts covering daily-life and payment-related scenarios. The synthetic speech is physically replayed through 9 loudspeakers and captured using 8 phones and 2 smart glasses. The dataset accompanies our work on ReplayFactor and supports research on cross-domain generalization.

The full dataset and protocols will be released after the paper review process is complete.

Dataset Details

Item	Details
Data scale	184,954 utterances; 232.5 hours
Language	Chinese
text	Daily-life and payment-related scenarios
Source speakers	1,921
TTS models(22)	F5-TTS, CosyVoice2, CosyVoice3, VITS_inhouse, Index-TTS1, GPTSoVITS, Fish-1.4, Llasa-1B, XVoice, FireRedTTS2, CosyVoice-300M, Index-TTS1.5, SparkTTS, FireRedTTS, Ming-OMNI-TTS, VITA-QinYu, Step_Audio_TTS, Llasa-3B, Index-TTS2, StyleSpeech, E2-TTS, OmniVoice
Play devices(9)	6 consumer loudspeakers: XiaoAI-speaker, SOAIY, JBL, HP-speaker, Sanag, Huawei-speaker; and 3 hight-fidelity loudspeakers: Edifier-MR4, Yamaha-HS5, and Genelec-8010.
Record devices: phones (8)	HUAWEI_Mate30Pro, Vivo_X100Pro, HUAWEI_Pura80, OPPO_FindX7Ultra, HUAWEI_Mate70Pro, Xiaomi-15Pro, iphone15, OPPO_FindX7Pro
Record devices: smart glasses (2)	Rokid_RG-glasses, Xiaomi_glass
Device pairs	90 playback–recording combinations
Playback distance	10, 30, and 50 cm

The source corpus comprises recordings from 1,921 speakers, while the texts cover daily-life and payment-related contents. We employ 22 state-of-art TTS models to generate deepfake speech, such as CosyVoice3 and LLaSA-1B, etc. To simulate practical spoofing scenarios, the generated speech is physically played through 9 loudspeakers, and re-recorded using 8 phones and 2 smart glasses, yielding 90 playback–recording device pairs.

Metadata and Protocols

Each utterance will include speaker ID, TTS model, playback device, recording device, distance (cm), and text. Protocol files will identify the recordings assigned to each split. File formats and field mappings will be documented upon release.

Split	Utterances	Speakers	TTS models	Playback devices	Recording devices
Train	77,063	1,921	13	5	7
Dev	2,932	1,437	5	5	7
Test	104,959	1,921	21	9	9
Overall	184,954	1,921	22	9	10

The release will include training and evaluation protocols, with out-of-domain settings covering unseen TTS models, unseen devices, and unseen model–device combinations. These protocols support evaluation of cross-domain generalization. Protocol file formats, split statistics will be documented upon release.

Date Structure

TBD

Access and Citation

Download instructions and the citation for the accompanying ReplayFactor paper will be added upon release.

Funded by: Ant Group

Contact: miao_he@zju.edu.cn, jiayizhou97@gmail.com