Release scope: benchmark only. This release includes RealCommBench and its paired Digital/test references. RealCommTrain and the digital train/dev partitions are not included. The benchmark package and non-commercial use terms are available to authorized repository users. The repository remains private pending opening of the approved-access release; public access requests are not yet open.

Dataset summary

RealComm pairs digital source utterances with recordings received through real mobile calls. Both bona fide and synthetic speech traverse the same recording conditions. This benchmark release supports audio deepfake detection, measurement of digital-to-call domain shift, and paired analyses of quality, spectra, and detector behavior.

The released test sources contain 560 utterances, balanced across bona fide/synthetic labels and English/Mandarin. Each source has 42 call versions. Synthetic speech comes from CosyVoice2, F5-TTS, FlexiVoice, IndexTTS2, MaskGCT, Vevo2, and Minimax.

Benchmark contents

Domain	Partition	Source utterances	Recordings
Digital	test	560	560
RealCommBench	test	560	23,520

The package contains 24,080 audio files from 560 source utterances. The two rows share the same underlying sources and must not be added as independent content. All files belong to the test partition.

The 42 test conditions combine over-the-air handset operation, over-the-air speakerphone operation, or line-to-microphone injection with directed device routes and available denoising settings. The labels off, on, and not_available are distinct.

Release policy

RealComm follows a phased release strategy. This release prioritizes standardized evaluation and reproduction of the digital-to-call benchmark results. RealCommTrain, including its development split, and the digital train/dev partitions are outside the scope of the current release. No release date is announced for those partitions.

The manuscript describes the full research corpus, including 2,800 digital sources and 12,320 RealCommTrain recordings. Its T and A+T results use the unreleased training data. This benchmark package alone does not enable reproduction of those adaptation training runs.

Access

Dataset files will require an approved individual Hugging Face account. The request form asks for affiliation and intended use, in addition to the account/contact information shared through Hugging Face. The maintainer reviews requests manually.

The private preparation repository is not yet accepting public access requests. The package contains one digital-reference ZIP and six call-audio ZIPs, together with a single test manifest and SHA-256 checksums. See file layout and evaluation instructions for the download commands and manifest schema; download commands apply after access opens.

Evaluation

Complete RealComm evaluation results will be published and maintained on SpoofRadar, our unified leaderboard for speech deepfake detection. The platform will present overall and condition-level results with model/checkpoint versions and evaluation settings. Zero-shot benchmark entries will be distinguished from adaptation results that use RealCommTrain. The manuscript reports a fixed experimental snapshot.

Use the frozen test manifest and report Digital/test and RealCommBench separately. The main metric is pooled equal error rate (EER) over the complete RealCommBench score set. Subgroup EERs are descriptive and should not be averaged to obtain the pooled result.

Retain the link from each call recording to its digital source. Both the digital and call files are reserved for evaluation; no training or development split is provided. Use independent data for checkpoint, threshold, and hyperparameter selection. Record the actual input duration, preprocessing, and score direction for each detector.

Label protocols prepared for a SpoofRadar inclusion request are available in protocols/spoofradar. The primary protocol covers the 23,520 call recordings; digital references have a separate protocol. Preparation of these files does not imply leaderboard acceptance.

Scope and limitations

The corpus covers the documented mobile-call collection settings and seven synthesis systems. Train and test use the same generator set; the main generalization questions concern held-out source content and communication paths. Multiple recordings of one source are repeated channel observations, not independent utterances. The dataset does not represent every device, service, or communication channel.

License and citation

Use is limited to non-commercial purposes; commercial use is prohibited. Access requires individual approval. See the RealComm Benchmark Non-commercial Use Terms. The final citation will be supplied when available. Refer to the manuscript draft for the current study; the extended draft is distinct from the earlier accepted SLT paper.

Downloads last month: 7

Total file size:

9.22 GB