AI Voice Models May Forget How to Mimic Specific Voices
▼ Summary
– AI companies use guardrails to prevent misuse, but clever techniques can sometimes bypass these restrictions to extract unwanted information.
– Machine unlearning aims to remove specific data from AI models entirely, creating a version that never learned the unwanted information.
– Jinju Kim compares guardrails to fences around bad data, while unlearning removes the data completely, leaving nothing to access.
– Text-to-speech models complicate unlearning, as they must forget trained voices and avoid mimicking specific untrained voices while maintaining performance.
– The research team demonstrated voice unlearning with Meta’s VoiceBox, showing a 75% reduction in mimicking forgotten voices but a slight performance drop for permitted voices.
AI voice models are facing a growing challenge: learning how to forget. While these systems excel at mimicking human speech patterns, researchers are now exploring ways to make them intentionally lose the ability to replicate specific voices. This emerging field, known as machine unlearning, could help address privacy concerns and prevent potential misuse of voice replication technology.
Most AI companies implement safeguards to prevent their models from generating harmful or unauthorized content. These protective measures act like digital barriers, stopping users from accessing restricted information through standard prompts. However, determined individuals often find ways to bypass these limitations, revealing how sensitive data remains embedded within the system’s architecture. Machine unlearning offers a more permanent solution by fundamentally altering the model’s knowledge base.
Jinju Kim, a researcher at Sungkyunkwan University, compares traditional safeguards to fences around prohibited content. “The barriers might stop casual attempts,” she explains, “but dedicated individuals will always find ways around them.” Unlearning technology takes a different approach by completely removing the problematic data, leaving nothing to circumvent.
Voice replication systems present unique challenges for unlearning techniques. Modern text-to-speech models can imitate voices they’ve never encountered in training data, given just a brief audio sample. This remarkable capability means unlearning must accomplish two things simultaneously: erase memory of trained voices and prevent imitation of new ones – all while maintaining performance for approved voices.
Kim’s team developed a method using a modified version of Meta’s VoiceBox model. Their approach trains the system to respond to prompts targeting redacted voices by generating random vocal patterns instead. The model essentially teaches itself to create new, artificial voices to replace the ones being forgotten. Testing showed the modified system reduced voice similarity by over 75% for forgotten speakers while only experiencing a minimal 2.8% performance drop for permitted voices.
The unlearning process isn’t instantaneous – it can take several days depending on how many voices need removal. Each targeted voice requires about five minutes of reference audio for the system to properly forget it. Researchers replace the original voice data with randomized patterns, making reconstruction of the forgotten voices virtually impossible.
Vaidehi Patil, a machine unlearning specialist at UNC Chapel Hill, notes the significance of this work. “While we’ve seen randomness optimization in other AI applications,” she observes, “this represents groundbreaking progress for speech systems.” The research team will present their findings at the International Conference on Machine Learning, offering new possibilities for responsible voice technology development.
(Source: Technology Review)
