Guarding Against Audio Risks: Introducing AudioSafetyBench and AudioGuard

Audio has swiftly become a key interface for foundation models, powering applications like voice assistants. But with this rise comes an array of complex safety challenges unique to audio. Real-world risks involve not just the spoken word but also harmful sound events, speaker attributes, and dangerous content combinations. Who would've thought that the sound of a child's voice mixed with inappropriate content could pose such a threat?

The Challenge of Ensuring Audio Safety

The nature of audio makes it inherently difficult to develop comprehensive benchmarks or guardrails. Unlike text, audio involves layers like speaker identity and sound events that can complicate risk assessments. Traditional methods fall short in addressing these complex audio risks.

Enter AudioSafetyBench, a novel policy-based audio safety benchmark. It approaches audio safety from a broad perspective, supporting multiple languages, suspicious voice types (think celebrity impersonations), and risky voice-content combinations. This isn't just about catching unsafe utterances. It's about understanding the intricate interplay of audio elements that could be harmful.

Introducing AudioGuard

To combat these nuanced threats, AudioGuard steps in with a dual approach. It features SoundGuard for detecting audio-native issues at the waveform level and ContentGuard for ensuring semantic protection based on policies. This two-pronged system is more than just an incremental improvement. It's essential for advancing how we think about and implement audio safety.

Extensive experiments reveal that AudioGuard consistently outperforms existing audio-LLM-based baselines. What stands out is its ability to do so with significantly lower latency. In an era where speed and accuracy are important, AudioGuard's performance is a game changer.

Why Audio Safety Matters

The key finding here's the proactive approach to audio safety. In a world where audio interfaces are ubiquitous, the potential for impersonation or misuse can't be ignored. The question isn't if these threats exist but how we'll address them. Will other developers follow this lead, or will they wait until a major incident forces their hand?

Code and data are available at AudioSafetyBench's repository, empowering researchers and developers to explore and address these risks. The ablation study reveals that while AudioGuard strengthens detection, there's still room for improvement. Continued research and iteration are essential.

Ultimately, AudioSafetyBench and AudioGuard represent significant strides in protecting audio systems. As audio interfaces continue to evolve, so too must our safety measures. In this domain, staying ahead of potential threats isn't just an option. It's a necessity.