🎤 MultiAPI-Spoof: Multi-API Audio Anti-Spoofing Dataset & Nes2Net-LA

📄 Paper on arXiv 🔗 Project Code 📥 Dataset Download on Google Drive 📂 Dataset Details & Audio Examples 💾 Pretrained Models (XLSR-Nes2Net-LA)

---

🌟 Overview

This repository contains code MultiAPI Spoof, a multi-API audio anti-spoofing dataset, and the Nes2Net-LA model, a local-attention enhanced anti-spoofing network.

Existing speech anti-spoofing datasets mostly rely on a limited set of public TTS/VC models, creating a gap from real-world scenarios where commercial systems use diverse, proprietary APIs. MultiAPI Spoof addresses this by including ~230 hours of synthetic speech from 30 APIs, covering commercial TTS services, open-source models, and online TTS platforms.

Two tasks are provided:

1. 🎯 Anti-Spoofing Detection: Classify bona fide vs. spoofed audio.
2. 🕵️‍♂️ API Tracing: Identify which API generated a given spoofed sample.

Nes2Net-LA improves Nes2Net-X by adding local attention, boosting local context modeling and fine-grained spoofing feature extraction.

---

⚙️ Installation

# Clone the repo
git clone https://github.com/XuepingZhang/MultiAPI-Spoof.git
cd MultiAPI-Spoof

# Create conda environment
conda create -n multiapi-spoof python=3.10
conda activate multiapi-spoof

#install fairseq
git clone https://github.com/facebookresearch/fairseq.git fairseq_dir
cd fairseq_dir
git checkout a54021305d6b3c
pip install --editable ./

# Install dependencies
pip install -r requirements.txt

---

🚀 Quick Start

The following examples are training scripts intended for cluster environments such as Slurm. Please modify the script according to your own environment and replace all paths with those specific to your setup. Pretrained XLSR

The pretrained model XLSR can be found at this link.

🟢 Anti-Spoofing Training

cd job/xlsr2_nes2net_ATT
sbatch sh/multi_train.slurm

🟢 API Tracing Training

cd job/api_tracing
sbatch sh/multi_train.slurm

🟢 Inference / Evaluation

cd job/xlsr2_nes2net_ATT
sbatch sh/multi_eval.slurm

---

📊 Experimental Results

Anti-Spoofing Performance

Model	ITW	MultiAPI Spoof	AI4T
XLSR+Nes2Net-LA	1.42 / 0.020 / 0.021	0.56 / 0.008 / 0.008	5.64 / 0.051 / 0.077

Metrics: EER / minDCF / actDCF

API Tracing Performance

Class	Precision	Recall	F1
Seen APIs (A0–A20)	0.950	0.923	0.936
Unseen APIs (A24–A29)	0.972	0.520	0.678
Overall	0.770	0.917	0.782

---

📖 Citation

If you use this code or dataset, please cite:

@misc{zhang2025multiapispoofmultiapidataset,
      title={MultiAPI Spoof: A Multi-API Dataset and Local-Attention Network for Speech Anti-spoofing Detection}, 
      author={Xueping Zhang and Zhenshan Zhang and Yechen Wang and Linxi Li and Liwei Jin and Ming Li},
      year={2025},
      eprint={2512.07352},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2512.07352}, 
}