PC-Mix Dataset

Partial Spoof Dataset with Controlled Speech-Background Mixing

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

Resources

GitHub

Code, protocols, and demo assets.

GitHub (TBD)

Replace “#” with your repo link.

Paper

Dataset description and baselines.

PDF / arXiv (TBD)

Replace “#” with your paper link.

Hugging Face

Dataset card and hosted files.

HF Dataset (TBD)

Replace “#” with your HF link.

Download

Direct download entry (mirror / drive / checksum).

Download (TBD)

Replace “#” with your download link.

Dataset Statistics

Split	Mixed Samples	Original Samples	Background Source	Event Source	Fusion Method
Train	25,380	6,345	SONYC	60% AudioLDM2 40% UrbanSound8K (0–1s)	Ducking Overlay
Dev	24,844	6,211	IHTApark-UBS	60% AudioLDM2 40% ESC-50 (0–1s)	Ducking Overlay

Evaluation Sets

Subset	Mixed	Original	Background	Event	Fusion
E0 Baseline	17,809	17,809	SONYC	60% AudioLDM2 40% FSD50K (0–1s)	Ducking Overlay
E1 Generator OOD	14,247	-	SONYC	AudioGen	Ducking Overlay
E2 Fusion OOD	14,247	-	SONYC	AudioLDM2	Energy Matching + Crossfade (20–80ms)
E3 Background OOD	17,809	-	DEMAND	AudioLDM2	Ducking Overlay
E4 Noise OOD	7,125	-	WHAM Noise	AudioLDM2	Ducking Overlay

Audio Demo

Example 1

Type	Audio
Original
Speech Bona + Background Bona
Speech Bona + Background Spoof
Speech Spoof + Background Bona
Speech Spoof + Background Spoof