RootAccess4Life/cond-dim-samples

cond-DIM — audio samples

Synthesized audio backing the paper "cond-DIM: One Method, Three Conditioning Paths for Host-Independent Speaker Unlearning in Autoregressive Zero-Shot TTS." ~2,450 clips: every backbone, every baseline, and the relearn stress test.

🔇 Prefer listening in the browser? → Live demo Space

The one thing to listen for

Each speaker appears as a pair:

clip	meaning
cB_…	baseline clone — the un-edited model cloning that speaker. This is the voice being copied.
cE_…	edited clone — the same model, same reference, after cond-DIM unlearning. The identity should be gone.
cRL_…	relearn clone — after a white-box attacker re-finetunes to recover the erased speaker.

Play cB then cE for the same speaker: intelligible speech in both, but the voice identity should no longer match. cRL asks whether an adversary can undo that.

Filename convention

{cB|cE|cRL}_{speaker}_{sentence}.wav

Speaker ids are VCTK (p225, p272, …) or LibriTTS (103, 118, …) depending on the split; {sentence} indexes a fixed sentence bank, so the same index is the same text across every directory. TruS-F5 clips carry an extra seed field (cB_s0_p225_0.wav).

Directory map

Main result — cond-DIM on three backbones

directory	backbone
xtts_clones/	XTTS-v2
tortoise_clones/	Tortoise-TTS
indextts_clones/	IndexTTS-1.5

Stress tests

directory	what it shows
xtts_relearn_clones/	white-box relearn attack on XTTS-v2 (cRL_)
sequential/{xtts,tortoise,indextts}_clones/	sequential opt-out — speakers erased one after another

Baselines (each in {method}_{backbone}_clones/)

additive_gaussian · gradient_ascent_to_forget · random_reference_substitution · task_arithmetic · zero_pooled_identity

TruS on F5-TTS

directory	contents
trus_f5_clones/	TruS-F5 clones (VCTK)
trus_f5_libritts_clones/	TruS-F5 clones (LibriTTS)
trus_f5_frontier/	α/threshold frontier sweeps (JSON metrics, not audio)

Loading

Stream a single clip without cloning the whole set:

from huggingface_hub import hf_hub_download

p = hf_hub_download(
    repo_id="RootAccess4Life/cond-dim-samples",
    filename="xtts_clones/cE_103_0.wav",
    repo_type="dataset",
)

Or grab one backbone:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="RootAccess4Life/cond-dim-samples",
    repo_type="dataset",
    allow_patterns=["xtts_clones/*"],
)

Provenance and licensing

Clips are synthesized by XTTS-v2, Tortoise-TTS, IndexTTS-1.5 and F5-TTS from reference audio in CSTR VCTK-Corpus-0.92 (CC-BY-4.0, DOI 10.7488/ds/2645) and LibriTTS (CC-BY-4.0, OpenSLR SLR60). This dataset is released CC-BY-4.0 to match those sources.

The generating models keep their own licenses — notably XTTS-v2 is non-commercial (CPML). Those terms constrain the models, not these audio files, but review them before building on the pipeline.

Intended use and limits

Released for research and demonstration: verifying the paper's claims by ear, and supporting work on speaker unlearning, privacy, and the right to be forgotten in speech synthesis.

These are synthetic clones of real speakers, produced to demonstrate removing voice identity. Do not use them to impersonate anyone, to train or evaluate systems that attribute speech to the original speakers, or as evidence any of these people said these words — they did not. The cB_ clips in particular are deliberate voice clones and should be handled accordingly.

Related

• 💻 Code: https://github.com/pujariaditya/cond-dim-tts-unlearning
• 🔇 Live demo: https://huggingface.co/spaces/RootAccess4Life/cond-dim-demo
• 📦 Mirrors: XTTS-v2 · IndexTTS-1.5

Citation

@misc{pujari_conddim,
  title  = {cond-DIM: One Method, Three Conditioning Paths for Host-Independent
            Speaker Unlearning in Autoregressive Zero-Shot TTS},
  author = {Pujari, Aditya and Rattani, Ajita},
  note   = {Preprint},
}