Exploratory Evaluation of Speech Content Masking
Jennifer Williams, Karla Pizzi, Paul-Gauthier Noe, Sneha Das
TL;DR
The paper addresses content-level privacy in speech by introducing content masking, a toy problem to redact selected words while preserving intelligibility of other content. It presents a baseline approach that masks content by altering sequences of VQ-VAE phone codes and re-synthesizes with WaveRNN, alongside perfect-masked natural-speech baselines for comparison. The authors evaluate the impact on ASR and ASV, showing that mask type and location meaningfully affect recognition and verification performance, with VQ-VAE-based masking causing substantial degradation in both tasks. This work demonstrates the feasibility and challenges of content masking for privacy-preserving speech and highlights directions for improving vocoders, recoverability, and attacker modeling to inform real-world data-sharing and privacy-preserving applications.
Abstract
Most recent speech privacy efforts have focused on anonymizing acoustic speaker attributes but there has not been as much research into protecting information from speech content. We introduce a toy problem that explores an emerging type of privacy called "content masking" which conceals selected words and phrases in speech. In our efforts to define this problem space, we evaluate an introductory baseline masking technique based on modifying sequences of discrete phone representations (phone codes) produced from a pre-trained vector-quantized variational autoencoder (VQ-VAE) and re-synthesized using WaveRNN. We investigate three different masking locations and three types of masking strategies: noise substitution, word deletion, and phone sequence reversal. Our work attempts to characterize how masking affects two downstream tasks: automatic speech recognition (ASR) and automatic speaker verification (ASV). We observe how the different masks types and locations impact these downstream tasks and discuss how these issues may influence privacy goals.