MultiAPI Spoof: Multi-Source Audio Anti-Spoofing Dataset
November 21, 2025
Introduction
MultiAPI-Spoof is a multi-source audio anti-spoofing dataset that contains approximately 230 hours of audio. It includes synthetic audio generated by commercial TTS services, open-source models, and Chinese TTS websites.
This dataset is designed to support research and model training for audio anti-spoofing.
- 📥 Download link: MultiAPI Spoof dataset
- 📄 Paper on arXiv (coming soon)
- 🖥️ Code on github
- License: For academic research only.
Spoofed Audio Samples
Data Sources
Our new dataset, MultiAPI Spoof, contains speech samples generated from a variety of API sources, including:
- Commercial TTS APIs – speech generated by commercial services.
- Open-Source Model Generation – speech generated by open-source models.
- TTS Websites – speech on TTS web platforms.
The dataset is organized into 30 API, labeled A0–A29, with each group corresponding to a unique speech generation API source. The duration of speech in each API ranges from 0.2 to 12 hours.
Dataset Split
| API NO. | train | dev | eval |
|---|---|---|---|
| A0–A20 | 70% | 10% | 20% |
| A21–A23 | / | 100% | / |
| A24–A29 | / | / | 100% |
Metadata
The dataset includes three metadata files: MultiAPI_train.txt, MultiAPI_dev.txt, and MultiAPI_eval.txt.
Each line has four fields:
| audio path | api | class_label |
|---|---|---|
| XXX.mp3 | A0 | spoofed |
| XXX.mp3 | - | bonafide |
Anti-spoofing Detection Demo
We trained a speech anti-spoofing model using our newly proposed MultiAPI Spoof dataset along with several public datasets. The model has been deployed online. 👉 Try it out through our interactive demo — simply upload an audio sample to get a spoofing score and classification result in real time.
Remark: This model is for reference only. As spoofing technologies evolve rapidly, detection results may contain errors. We will regularly update our model. The model and dataset are provided solely for academic research and are strictly prohibited from commercial use.