eliya/forensics_0.3B_v2_deepfake_age_gender_classifier
Audio Classification • Updated • 2 • 2
Datasets:
Acknowledge the license to access this dataset
Requests are reviewed manually.
This dataset is released under CC-BY-NC-4.0 — free for personal and research use. Commercial use is prohibited under this license; for a commercial license, contact eliya@vocos.io.
Log in or Sign Up to review the conditions and access this dataset content.
qwen3_deepfake_dataset
515.5 hours of synthetic speech deepfakes — 364,640 clips across 9 corpora and 6 languages (Mandarin, English, Spanish, French, Italian, Japanese), generated via voice cloning and hundreds of designed voice profiles.
Generated with Qwen3-TTS, built to train/evaluate speech deepfake detectors (part of the Forensics model family).
Built to improve detection of voice cloning, deepfakes, highly realistic synthetic voices, voice conversion/changing, and similar identity-manipulation techniques.
How it was built
Real utterances from 9 public speech corpora (Aishell, Aishell3, CommonVoice — en/es/fr/it/ja, LibriTTS, VCTK-Corpus) were fed through Qwen3-TTS in three modes:
- clone_self — resynthesize a speaker's own line using their own voice as the reference
- clone_cross — resynthesize a speaker's line using a different donor speaker's voice as the reference (identity-swap deepfake)
- design — synthesize speech from a designed target voice profile (age/gender/pitch), not cloned from any specific reference
Every generated clip's speaker age and gender were predicted and attached as columns. The metadata additionally includes one bonafide row per unique real source utterance, for reference — the audio for those rows is not included in this repo (see below).
Example: one source recording, every variant generated from it
Since this dataset is gated, here's exactly what you're requesting access to — a real example, not a made-up one. One Aishell recording (BAC009S0002W0135) and everything generated from it:
| File | Method | What it is |
|---|---|---|
speechfake/Real/Aishell/train/S0002/BAC009S0002W0135.wav |
(original — not bundled, see above) | The real recording everything below is derived from |
Aishell/BAC009S0002W0135/BAC009S0002W0135_clone_self.wav |
clone_self | Same content, resynthesized in the same speaker's own voice |
Aishell/BAC009S0002W0135/BAC009S0002W0135_clone_cross.wav |
clone_cross | Same content, resynthesized in a different donor speaker's voice — an identity-swap deepfake |
..._designed_f_elder70s_lowpitched_v_379.wav |
design | Same content, synthesized from an invented voice profile (female, elderly, low-pitched) — no real reference voice at all |
..._designed_m_early60s_mediumlow_vo_327.wav |
design | Same content, a different invented profile (male, early 60s, medium-low pitch) |
..._designed_m_late60s_medium_voice_012.wav |
design | Another invented profile (male, late 60s, medium pitch) |
..._designed_m_mid50s_medium_voice_057.wav |
design | Another invented profile (male, mid-50s, medium pitch) |
design draws from hundreds of predefined voice-profile descriptions (age bracket × gender × pitch register) rather than cloning any real person — the same source line gets resynthesized in dozens of different invented voices, as shown above.
What's in this repo
- The generated fake audio (
.wav), organized as<source_dataset>/<sample_id>/<file>.wav. metadata/metadata_age_gender_train.csvandmetadata/metadata_age_gender_test.csv— one row per clip.
What's not in this repo
Rows labeled bonafide describe the original real recordings this dataset was built from. Their new_file/orig_file/donor_file columns point at paths like speechfake/Real/Aishell/train/S0002/BAC009S0002W0135.wav — that audio lives in its own original public dataset (Aishell, Aishell3, CommonVoice, LibriTTS, or VCTK-Corpus, matching the dataset column) and is not duplicated here. Fetch it from the original source if you need the real counterpart.
Columns
| column | meaning |
|---|---|
new_file / file |
path to this row's audio (relative to repo root for fake rows; a reference-only path for bonafide rows — see above) |
orig_file |
the real utterance this clip was cloned from (reference-only path) |
donor_file |
for clone_cross, the donor speaker's reference utterance (reference-only path) |
duration |
seconds |
method |
qwen_clone_self, qwen_clone_cross, qwen_design, or real |
label |
fake or bonafide |
dataset |
source corpus |
predicted_age / predicted_gender |
model-predicted speaker attributes |
speaker / sample_id |
source speaker/utterance identifier |
Use
This dataset is gated — request access on this page, then authenticate before downloading:
huggingface-cli login # or: export HF_TOKEN=hf_xxx
Load the audio directly (organized as 9 configs, one per source corpus):
from datasets import load_dataset, get_dataset_config_names
get_dataset_config_names("eliya/qwen3_deepfake_dataset") # list the 9 corpora
ds = load_dataset("eliya/qwen3_deepfake_dataset", "Aishell", split="train")
sample = ds[0]
sample["audio"]["array"], sample["audio"]["sampling_rate"] # decoded waveform + rate
# For the larger corpora, stream instead of downloading everything upfront:
ds = load_dataset("eliya/qwen3_deepfake_dataset", "CommonVoice", split="train", streaming=True)
next(iter(ds))
Or work from the full reference metadata directly (all 9 corpora, includes bonafide provenance rows):
import pandas as pd
df = pd.read_csv("metadata/metadata_age_gender_train.csv")
fake_only = df[df["label"] == "fake"] # bonafide rows have no bundled audio
License
CC-BY-NC-4.0 — free for personal and research use. For commercial use, contact eliya@vocos.io.
- Downloads last month
- 55
Models trained or fine-tuned on eliya/qwen3_deepfake_dataset