Explaining Speaker and Spoof Embeddings via Probing

TL;DR

This paper explores how well deep neural network-based spoof embeddings retain speaker-related information, revealing that they preserve key traits like gender and speaking rate, which impacts their explainability in spoof detection.

Contribution

It extends speaker embedding explainability to spoof embeddings, analyzing the extent to which these embeddings encode speaker traits and their implications for model robustness.

Findings

Spoof embeddings retain traits like gender and speaking rate.

Spoof detectors partially preserve speaker traits for robustness.

Experiments conducted on ASVspoof 2019 LA dataset.

Abstract

This study investigates the explainability of embedding representations, specifically those used in modern audio spoofing detection systems based on deep neural networks, known as spoof embeddings. Building on established work in speaker embedding explainability, we examine how well these spoof embeddings capture speaker-related information. We train simple neural classifiers using either speaker or spoof embeddings as input, with speaker-related attributes as target labels. These attributes are categorized into two groups: metadata-based traits (e.g., gender, age) and acoustic traits (e.g., fundamental frequency, speaking rate). Our experiments on the ASVspoof 2019 LA evaluation set demonstrate that spoof embeddings preserve several key traits, including gender, speaking rate, F0, and duration. Further analysis of gender and speaking rate indicates that the spoofing detector partially preserves these traits, potentially to ensure the decision process remains robust against them.