"Alexa, can you forget me?" Machine Unlearning Benchmark in Spoken Language Understanding

Right to be forgotten is technically challenging in speech AI: The paper demonstrates that removing a specific speaker's data from voice assistants like Alexa is significantly more difficult than simply deleting files, requiring specialized "unlearning" techniques that must balance completely removing the speaker's influence while maintaining the system's overall performance.

No single unlearning method works best across all situations: The study evaluated eight different unlearning techniques across four languages and found "significant differences in the effectiveness and computational feasibility of various techniques," suggesting that organizations may need to adopt different approaches depending on their specific AI systems and use cases.

Current unlearning methods have real trade-offs: The researchers developed a new metric to measure three competing priorities simultaneously—how completely data is removed (efficacy), whether the AI still works well for other users (utility), and how much computing power is required (efficiency)—revealing that improving one often comes at the cost of another.

This is the first comprehensive benchmark for voice AI unlearning: UnSLU-BENCH is the first standardized test for machine unlearning in spoken language understanding, covering four datasets across four languages, establishing a foundation for the field and enabling fair comparison of different unlearning approaches going forward.

Responsible AI requires new technical infrastructure: The paper highlights that supporting privacy rights like "the right to be forgotten" in modern AI systems demands proactive technical solutions built into systems from the start, not just reactive data deletion policies.