Phoneme Discretized Saliency Maps for Explainable Detection of AI-Generated Voice

TL;DR

This work addresses the lack of interpretable explanations for AI-generated voice detectors by introducing Phoneme Discretized Saliency Maps (PDSM), which discretize saliency maps using phoneme posteriorgrams (PPGs) derived from ASR-based phoneme boundaries. The method couples posthoc attributions (e.g., Integrated Gradients, GradSHAP) with phoneme-level pooling to produce binary masks that retain the most informative phonemes, improving both faithfulness and understandability. Empirically, PDSM outperforms standard saliency maps across Tacotron2- and FastSpeech2-generated speech detection tasks and provides explanations aligned with phoneme structure; the authors also release the dataset used. The approach has practical impact for building transparent detectors of AI-generated speech and may generalize to other speech-analysis problems such as emotion recognition or clinical speech applications, facilitating better trust and debugging in real-world deployments. $FF_n = f(X_n)_c - f(X_n \odot (1-M))_c$ is used to quantify faithfulness, while $PPG$ provides phoneme boundaries that anchor the explanations.$

Abstract

In this paper, we propose Phoneme Discretized Saliency Maps (PDSM), a discretization algorithm for saliency maps that takes advantage of phoneme boundaries for explainable detection of AI-generated voice. We experimentally show with two different Text-to-Speech systems (i.e., Tacotron2 and Fastspeech2) that the proposed algorithm produces saliency maps that result in more faithful explanations compared to standard posthoc explanation methods. Moreover, by associating the saliency maps to the phoneme representations, this methodology generates explanations that tend to be more understandable than standard saliency maps on magnitude spectrograms.

Figures (6)

• Figure 1: (left) Input Spectrogram, (right) Interpretations obtained for a trained network with the GradSHAP algorithm. We overlay the boundaries of the time-limited Gaussian noise on top of the interpretations.
• Figure 2: (left) Input Spectrogram for an AI-generated voice, (right) GradSHAP interpretations obtained for a CNN14 trained network
• Figure 3: Steps of the PDSM algorithm.
• Figure 4: (left-most, left-center) Analysis of faithfulness with respect to different numbers of phonemes retained in the binary mask for Tacotron2 and FastSpeech2 respectively. (right-center, right) Same analysis of faithfulness normalized over the length of audio sample, thus representing fraction of audio retained for the binary mask.
• Figure 5: Global Phoneme Importances Obtained by Tacotron2 and FastSpeech2. Note that $<>$ indicates silence.