SafeEar

SafeEar is an open-source AI Framework developed jointly by Zhejiang University and Tsinghua university to address the growing threat of audio deepfakes. Unlike conventional detection tools that rely on full speech reconstruction or language-specific models, SafeEar employs a novel neural audio codec-based architecture that deliberately decouples semantic meaning from raw acoustic patterns. This design enables highly accurate forgery detection without exposing sensitive spoken content—making it especially valuable for privacy-critical APPlications such as legal forensics, financial authentication, and national SECurity monitoring.

What it does

SafeEar analyzes audio inputs to determine whether they have been synthetically generated or manipulated using Voice Cloning, TTS, or other generative techniques. Its core innovation lies in operating exclusively on low-level acoustic representations—such as pitch contours, spectral envelopes, and temporal dynamics—while discarding phonemic or lexical Information during inference. This ensures that no Intelligible speech content is processed or retAIned, significantly reducing privacy risks associated with audio analysis.

Key capabilities

• Privacy-Preserving detection: Uses a disentangled encoder-decoder structure to isolate and discard semantic content before analysis.
• multilingual robustness: Trained and validated on CvoiceFake, a curated dataset of 1.5 million multilingual audio Samples spanning English, Mandarin, Japanese, Korean, Russian, and more.
• High detection accuracy: Achieves an Equal Error Rate (EER) of just 2.02% on Standard benchmarks—including ASVspoof 2019 and FakeAVCeleb—outperfORMing many state-of-the-art alternatives.
• Adversarial Resilience: Designed to withstand content recovery attacks and channel distortions common in real-world deployments like phone calls or streaming platforms.
• Real-world Generalization: Incorporates simulated environmental noise, bandwidth liMITations, and transmission artifACTs to improve performance under imperfect conditions.

Who it is for

SafeEar is built for researchers, forensic analysts, cybersecurity Professionals, and developers working on trust and integrity in voice-driven systems. It’s particularly suited for organizations needing auditable, explainable, and privacy-compliant audio verification—such as law enforcement agencies validating evIDEnce, banks securing voice-authenticated transactions, or social media platforms moderating misinformation.

How to get started

The framework is publicly available on GitHub with documentation, training scripts, pretrained models, and the CVoiceFake dataset. Users can run Inference locally or integrate the model into existing pipelines via Python APIs. No registration or cloud account is required; however, familiarity with PyTorch and audio preprocessing is recommended for customization.

Pricing and access caveat

SafeEar is fully open-source and free to use under permissive licensing. There are no subscription tiers, usage limits, or commercial restrictions. However, deployment requires local compute resources—no hosted API or web interface is officially provided. Users should verify compliance with local data governance lAWS when applying SafeEar to regulated audio data.