PC-Mix Dataset

Partial Spoof Dataset with Controlled Speech–Environmental Sounds Mixing

PC-Mix pairs speech from PartialSpoof v1.2 with a self-curated partial-spoof environmental sounds pool to create controlled speech–environment authenticity combinations. The original samples are collected from VGGSound and SBCSAE.

Resources

Resource	Description	Link
GitHub	Code, protocols, and preprocessing scripts.	Anonymous GitHub
Paper	Dataset description and baselines.	PDF / arXiv (TBD)
Hugging Face	Dataset card and hosted files.	HF Dataset
Download	Direct access to the packaged dataset files.	Download Files

Download and Setup

PC-Mix is hosted on Hugging Face under Alphawarheads/PC-Mix. The dataset directories are compressed with tar and zstd, then split into fixed-size volumes for more convenient downloading. All archive parts belonging to the same directory must be downloaded before extraction.

Option 1: Download with Hugging Face CLI

pip install -U "huggingface_hub[cli]"

huggingface-cli download Alphawarheads/PC-Mix \
  --repo-type dataset \
  --local-dir ./PC-Mix

Option 2: Clone with Git LFS

git lfs install

git clone https://huggingface.co/datasets/Alphawarheads/PC-Mix

Option 3: Access from Python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Alphawarheads/PC-Mix",
    repo_type="dataset",
    local_dir="./PC-Mix",
)

Install the extraction dependency

The archive volumes use Zstandard compression.

Using Conda:

conda install -c conda-forge zstd

Using Ubuntu/Debian:

sudo apt install zstd

Packaged File Structure

The files hosted on Hugging Face are organized as follows:

PC-Mix/
├── README.md
├── buildSBCori.py
├── extract_orig_from_mixed_protocol.py
├── FILELIST.txt
├── SHA256SUMS
│
├── Mixed_soundsV1.2/
│   ├── Mixed_soundsV1.2.tar.zst.part-000
│   ├── Mixed_soundsV1.2.tar.zst.part-001
│   ├── Mixed_soundsV1.2.tar.zst.part-002
│   └── ...
│
└── Partialspoof_background/
    ├── Partialspoof_background.tar.zst.part-000
    ├── Partialspoof_background.tar.zst.part-001
    └── ...

The split volumes preserve the original directory names and all internal relative paths. After extraction, the dataset returns to the directory structure described in Dataset Structure.

Verify downloaded volumes

Run the checksum verification from the downloaded repository root:

cd ./PC-Mix
sha256sum -c SHA256SUMS

Every successfully downloaded volume should be reported as OK.

Extract the Split Archives

Create a destination directory:

mkdir -p ./PC_Mix_extracted

Extract Mixed_soundsV1.2

cat ./PC-Mix/Mixed_soundsV1.2/Mixed_soundsV1.2.tar.zst.part-* \
  | zstd -dc \
  | tar -xf - -C ./PC_Mix_extracted

Extract Partialspoof_background

cat ./PC-Mix/Partialspoof_background/Partialspoof_background.tar.zst.part-* \
  | zstd -dc \
  | tar -xf - -C ./PC_Mix_extracted

After extraction:

PC_Mix_extracted/
├── Mixed_soundsV1.2/
└── Partialspoof_background/

Do not rename, individually decompress, or extract the .part-* files. They are consecutive parts of a single compressed archive and must be concatenated in numerical order.

Reconstruct the SBCSAE-Derived Original WAV Files

Due to the CC BY-ND 3.0 US license of SBCSAE, SBCSAE-derived cropped audio is not directly redistributed in PC-Mix. The repository provides the protocols and preprocessing scripts required to reconstruct these files locally after obtaining the official SBCSAE WAV data.

The reconstruction workflow uses:

• extract_orig_from_mixed_protocol.py
• buildSBCori.py
• the extracted Mixed_soundsV1.2 directory
• the official SBCSAE WAV directory

1. Remove SBCSAE-derived original files before redistribution

This command identifies SBCSAE-derived original entries through the mixed-data protocols and deletes their corresponding train_orig_* or eval_orig_* WAV files.

python extract_orig_from_mixed_protocol.py \
  --mixed_sounds_dir /dkucc/home/zz324/Audiodata/PC_Mix/Mixed_soundsV1.2 \
  --delete

This step is primarily used when preparing a redistributable release. Users downloading the public PC-Mix package normally receive a version in which these restricted files have already been removed.

2. Reconstruct the omitted WAV files locally

After downloading the official SBCSAE audio and placing its WAV files under a local WAV directory, run:

python buildSBCori.py \
  --input_dir /dkucc/home/zz324/Audiodata/WAV \
  --mixed_sounds_dir /dkucc/home/zz324/Audiodata/PC_Mix/Mixed_soundsV1.2 \
  --protocol_ref /dkucc/home/zz324/Audiodata/PC_Mix/Mixed_soundsV1.2 \
  --overwrite

Arguments:

• --input_dir: directory containing the officially obtained SBCSAE WAV files.
• --mixed_sounds_dir: extracted Mixed_soundsV1.2 directory in which the reconstructed original files will be written.
• --protocol_ref: directory containing the train and evaluation protocols used to locate the required SBCSAE source segments.
• --overwrite: regenerate target files even when files with the same names already exist.

The script reconstructs the omitted SBCSAE-derived original clips while preserving the expected PC-Mix file names and directory layout.

Audio Class Description

PC-Mix provides both utterance-level and frame-level annotations.

At the utterance level, each sample is assigned to one of five classes: one class for original real-world recordings and four classes for constructed mixtures with different speech/environmental-sound authenticity combinations.

At the frame level, PC-Mix focuses on constructed mixtures only. Each frame is assigned to one of four classes based on whether the speech component, the environmental-sound component, or both components are spoofed at that time.

Utterance-Level Classes

Class	Name	Speech	Environmental Sounds	Description
0	Original	—	—	A naturally recorded audio clip containing speech and environmental sounds, without artificial speech–environment mixing.
1	Bona fide + Bona fide	Bona fide	Bona fide	A constructed mixture where both the speech component and the environmental-sound component are bona fide.
2	Spoofed + Bona fide	Spoofed	Bona fide	A constructed mixture containing partially spoofed speech while the environmental-sound component remains bona fide.
3	Bona fide + Spoofed	Bona fide	Spoofed	A constructed mixture containing bona fide speech with partially spoofed environmental sounds.
4	Spoofed + Spoofed	Spoofed	Spoofed	A constructed mixture where both speech and environmental sounds contain partially spoofed regions.

Frame-Level Classes

Class	Name	Speech	Environmental Sounds	Description
0	Bona fide frame	Bona fide	Bona fide	Neither component is spoofed in this frame.
1	Speech-spoofed frame	Spoofed	Bona fide	The local spoofing occurs in the speech component only.
2	Environmental-sound-spoofed frame	Bona fide	Spoofed	The local spoofing occurs in the environmental-sound component only.
3	Both-spoofed frame	Spoofed	Spoofed	Both speech and environmental sounds are spoofed in this frame.

Dataset Structure

After the split archives have been extracted and the omitted SBCSAE-derived original clips have been reconstructed, the dataset follows this structure:

PC_Mix
├── README.md
│
├── Mixed_soundsV1.2                                  # mixed speech + environmental sounds audio
│   │
│   ├── train                                         # training split
│   │   ├── protocol.txt                              # mapping from mixed audio to source speech/environmental sounds
│   │   ├── wav                                       # mixed audio files, e.g., train_mix_0000000.wav
│   │   └── labels                                    # NumPy label files
│   │       ├── speech                                # speech-component segment labels
│   │       ├── env                                   # environmental-sounds component segment labels
│   │       ├── mix                                   # mixture-level segment labels
│   │       └── original_vs_others_seglab             # original-vs-non-original segment labels
│   │
│   ├── eval                                          # evaluation split
│   │   ├── protocol.txt                              # mapping from mixed audio to source speech/environmental sounds
│   │   ├── wav                                       # mixed audio files, e.g., eval_E0_mix_0000006.wav
│   │   └── labels                                    # NumPy label files
│   │       ├── speech
│   │       ├── env
│   │       ├── mix
│   │       └── original_vs_others_seglab
│   │
│   ├── label_PartialSpoof_train                      # text-format PartialSpoof-style labels for training
│   │   ├── wav-train.scp
│   │   ├── original_vs_others_utt_labels.txt
│   │   ├── speech                                    # train_protocol_*.txt
│   │   ├── env                                       # train_protocol_*.txt
│   │   ├── mix                                       # train_protocol_*.txt
│   │   └── original_vs_others_seglab                 # train_protocol_*.txt
│   │
│   └── label_PartialSpoof_eval                       # text-format PartialSpoof-style labels for evaluation
│       ├── wav-eval.scp
│       ├── original_vs_others_utt_labels.txt
│       ├── speech                                    # eval_protocol_*.txt
│       ├── env                                       # eval_protocol_*.txt
│       ├── mix                                       # eval_protocol_*.txt
│       └── original_vs_others_seglab                 # eval_protocol_*.txt
│
└── Partialspoof_background                           # source environmental sounds data
    │
    ├── train                                         # training environmental sounds split
    │   ├── protocol.txt                              # environmental sounds audio protocol
    │   ├── wav                                       # environmental sounds audio files
    │   └── labels                                    # segment labels in .npy/.txt formats
    │
    ├── eval                                          # evaluation environmental sounds split
    │   ├── protocol.txt                              # environmental sounds audio protocol
    │   ├── wav                                       # environmental sounds audio files
    │   └── labels                                    # segment labels in .npy/.txt formats
    │
    ├── label_PartialSpoof_train                      # text-format environmental sounds labels for training
    │   └── train_protocol_*.txt
    │
    └── label_PartialSpoof_eval                       # text-format environmental sounds labels for evaluation
        └── eval_protocol_*.txt

Audio Source

Component	Source	Description
Speech	PartialSpoof v1.2	Speech samples with partially spoofed speech annotations.
Environmental Sounds	Self-curated partial-spoof environmental sounds pool	Environmental sounds with controlled bona fide/spoofed component labels.
Original	VGGSound and SBCSAE	Original recordings without artificial speech–environment mixing.

Data Splits

Training Set

Split	Mixed Samples	Original Samples	Environmental Sounds Source	Event Source	Fusion Method
Train	25,380	12,155	SONYC	60% AudioLDM2; 40% UrbanSound8K (0–1 s)	Ducking Overlay

Evaluation Sets

Subset	Mixed	Original	Environmental Sounds	Event	Fusion
E0 Baseline	17,812	17,809	SONYC	60% AudioLDM2; 40% FSD50K (0–1 s)	Ducking Overlay
E1 Generator OOD	14,248	—	SONYC	AudioGen	Ducking Overlay
E2 Fusion OOD	14,248	—	SONYC	AudioLDM2	Energy Matching + Crossfade (20–80 ms)
E3 Environmental Sounds OOD	17,809	—	DEMAND	AudioLDM2	Ducking Overlay
E4 Noise OOD	7,125	—	WHAM Noise	AudioLDM2	Ducking Overlay

Metadata

Each mixed sample is associated with a protocol entry that records its mixed-audio ID, speech source, environmental sounds source, mixing duration, SNR, gain/scaling information, and offset information.

Example:

train_mix_0000000 mix CON_T_0000000.wav train_29_006663_s2.wav 1.000 1.400 snr_db=8.000 bg_scale=0.425705 offset_sec=0.909625 offset_samp=14554

Citation

License 🔐

PC-Mix is released for non-commercial research purposes. Part of this dataset is a derived dataset constructed by combining, mixing, cropping, and generating audio samples from multiple publicly available datasets and audio generation models. Users must comply with the original licenses of all upstream datasets and models.

• The PartialSpoof, VGGSound, and SONYC-UST datasets are released under the CC BY 4.0 license.

• The UrbanSound8K and WHAM! noise data are released under the CC BY-NC 4.0 license.

• The FSD50K dataset contains audio clips under multiple Creative Commons licenses, including CC0, CC BY, CC BY-NC, and CC Sampling+. Users should follow the per-clip license metadata provided by FSD50K.

• The DEMAND dataset is released under the CC BY-SA 3.0 license.

• The synthetic environmental event sounds generated by AudioLDM2 are produced using model weights released under the CC BY-NC-SA 4.0 license.

• The synthetic environmental event sounds generated by AudioGen are produced using model weights released under the CC BY-NC 4.0 license.

• The SBCSAE / Santa Barbara Corpus of Spoken American English is released under the CC BY-ND 3.0 US license. Due to the NoDerivatives condition, SBCSAE-derived audio clips are not directly redistributed in PC-Mix. We only provide the official download link, segment-level metadata, and preprocessing scripts so that users who have obtained SBCSAE from the official source can reproduce this subset locally.

Important: PC-Mix does not override the licenses of the original datasets or models. By using PC-Mix, users agree to comply with all upstream license terms.

Contact Information