You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This bundle contains speech synthesized from the voices of LibriSpeech (LibriVox) readers and channel-perturbed derivatives of LibriSpeech. By requesting access you agree to use it only for research on speech deepfake detection and source tracing, not to impersonate anyone, and to respect the per-source terms in DATA_LICENSES.md (F5-TTS, XTTS-v2 and EnCodec outputs are non-commercial research use only).

Log in or Sign Up to review the conditions and access this dataset content.

Where Modern Zero-Shot TTS Leaves Its Trace — data bundle

Audio, embeddings, manifests, result tables and provenance for the manuscript "Where Modern Zero-Shot TTS Leaves Its Trace: Localizing Source-Tracing Cues in Self-Supervised Speech Representations" (M. Lee, J. Choo, KAIST AI; under review, submitted to ICASSP 2027). Code and the claim-to-file map: https://github.com/Firminwoo1015/tts-source-tracing.

archive contents size
manifests_and_results.tar manifests, QC list, speaker folds, every result table of the paper (results/paper5c17), analysis_plan.md, provenance (provenance.json, per-file provenance_files.jsonl.gz), licenses 5.3 MB
embeddings.tar layer-wise mean-pooled SSL embeddings (WavLM-Large, HuBERT-Large, XLS-R, wav2vec2-LV60, w2v-BERT 2.0) for every condition used in the paper: clean, interventions, decoder-only, the seven perturbations, trim+RMS, controlled seeds 8.26 GB
audio_generated.tar TTS outputs of F5-TTS v1, XTTS-v2, Fun-CosyVoice3, Chatterbox, IndexTTS-1.5 (24 kHz), the analysis–synthesis interventions, and the decoder-only vocoder outputs 2.28 GB
audio_perturbed.tar the seven perturbations of the five systems and of LibriSpeech real (16 kHz; three realizations for noise and phase randomization) 4.45 GB
audio_subsets.tar controlled-seed subsets (seeds 101–103, 20 speakers × 3 utterances × 5 systems) 0.31 GB

Verify with sha256sum -c SHA256SUMS. Restore with tar -xf <archive>.tar from the root of the code repository.

Licenses

Per-source terms are in DATA_LICENSES.md (inside manifests_and_results.tar and in the code repository). LibriSpeech/VCTK derivatives: CC BY 4.0 with attribution. Outputs of Apache-2.0/MIT models (CosyVoice3, Chatterbox, IndexTTS-1.5, Vocos, BigVGAN-v2, DAC, ParallelWaveGAN): research use with attribution. Outputs of F5-TTS (CC BY-NC 4.0), XTTS-v2 (CPML) and EnCodec (weights CC BY-NC 4.0): non-commercial research use only. Voices are clones of public read-speech corpora; do not use them to impersonate anyone. Original LibriSpeech/VCTK files are not redistributed.

Reproduction

The main run follows each wrapper's native RNG policy, so generated audio cannot be regenerated bit-exactly; the released embeddings reproduce every embedding-based number via bash src/run_all_analyses.sh in the code repository, and src/gen/*.py regenerate audio for any system whose outputs are not redistributed.

Downloads last month
25