Speech Unlearning

TL;DR

This paper pioneers the study of machine unlearning in speech models, addressing privacy, bias, and data removal challenges by defining tasks and demonstrating the complexity of unlearning speech data.

Contribution

It introduces the concept of speech unlearning, defines fundamental tasks, and highlights the unique challenges compared to vision and NLP domains.

Findings

Unlearning speech data is more challenging than image or text data.

Defined sample and class unlearning tasks for speech models.

Identified key future research directions in speech unlearning.

Abstract

We introduce machine unlearning for speech tasks, a novel and underexplored research problem that aims to efficiently and effectively remove the influence of specific data from trained speech models without full retraining. This has important applications in privacy preservation, removal of outdated or noisy data, and bias mitigation. While machine unlearning has been studied in computer vision and natural language processing, its application to speech is largely unexplored due to the high-dimensional, sequential, and speaker-dependent nature of speech data. We define two fundamental speech unlearning tasks: sample unlearning, which removes individual data points (e.g., a voice recording), and class unlearning, which removes an entire category (e.g., all data from a speaker), while preserving performance on the remaining data. Experiments on keyword spotting and speaker identification demonstrate that unlearning speech data is significantly more challenging than unlearning image or text data. We conclude with key future directions in this area, including structured training, robust evaluation, feature-level unlearning, broader applications, scalable methods, and adversarial robustness.