Audio-XAI: Explainable Audio AI

• Audio-XAI is the application of explainable AI to audio models, uncovering which parts of acoustic signals contribute to predictions.
• It adapts methods from vision XAI—such as LRP, Grad-CAM, and perturbation techniques—to suit audio’s unique temporal and spectral characteristics.
• Empirical studies show that modality-specific explanations enhance trust and provide actionable insights in tasks like speech recognition and anomaly detection.

Audio-XAI denotes the application of explainable artificial intelligence to audio models, with the aim of identifying which parts of an acoustic signal, intermediate representation, or conditioning input contribute to a model’s output. In the literature represented here, Audio-XAI spans spoken-digit and speaker analysis, audio event classification, automatic speech recognition, anomalous sound detection, audio deepfake detection, cough analysis, neural acoustic embeddings, and text-to-audio generation. The field includes both post-hoc attribution methods and model-intrinsic approaches, and it addresses explanation targets at multiple granularities: waveform samples, time-frequency bins, phoneme-aligned segments, token-to-frame alignments, latent channels, and text-token influences on generated audio (Becker et al., 2018, Frommholz et al., 2023, Wu et al., 2023, Dumpala et al., 2024, Grinberg et al., 23 Jan 2025, Kang et al., 1 Feb 2025, Buck et al., 26 Jan 2026, Kawa et al., 11 May 2026, Kitłowski et al., 12 Jun 2026, Ranjan et al., 12 Jun 2026, Amado-Caballero et al., 20 Aug 2025).

1. Scope and development of the field

An early benchmark for Audio-XAI was introduced by "AudioMNIST: Exploring Explainable Artificial Intelligence for Audio Analysis on a Simple Benchmark" (Becker et al., 2018). That work coupled an open dataset of 30,000 recordings of spoken English digits with Layer-wise Relevance Propagation (LRP), systematic input manipulations, and audible heatmaps. It established a template in which explanations are not merely visual overlays but testable hypotheses about model feature selection.

Subsequent work broadened both the application range and the methodological repertoire. In audio event classification, LRP and DFT-LRP were used to compare waveform and spectrogram models and to expose representation-dependent decision strategies (Frommholz et al., 2023). In ASR, model-agnostic perturbation methods were adapted to phoneme recognition with explicit ground truth from TIMIT, and later transformer-specific token-to-frame explanations were proposed for Whisper and speech-augmented decoder-only models (Wu et al., 2023, Ranjan et al., 12 Jun 2026). Audio deepfake detection became a major focus, with Grinberg et al. introducing a waveform-level transformer relevancy method and large-scale analyses of speech, non-speech, phonetic content, and voice onsets or offsets (Grinberg et al., 23 Jan 2025). Parallel developments addressed industrial anomalous sound detection through band-level perturbation faithfulness tests (Buck et al., 26 Jan 2026), clinically oriented cough analysis through occlusion-weighted spectral features (Amado-Caballero et al., 20 Aug 2025), and text-to-audio generation through token-level factual and counterfactual masking (Kang et al., 1 Feb 2025).

A distinct line of work has attempted to make the representations themselves explainable. XANE learns acoustic embeddings whose dimensions are tied to 14 acoustic parameters, including reverberation, noise, overlap, and CODEC properties (Dumpala et al., 2024). APEX instead keeps a pretrained classifier frozen and inserts an invertible disentanglement layer to obtain audio-centric prototype explanations while preserving the original logits exactly (Kawa et al., 11 May 2026). This suggests that Audio-XAI is not limited to retrospective saliency estimation; it also includes representation design and architecture-level constraints intended to make explanations structurally meaningful.

2. Signal representations and explanation targets

Audio-XAI operates on a wider range of representations than image XAI. At the input level, models may consume raw waveforms, magnitude spectrograms, log-mel spectrograms, or latent acoustic embeddings. For waveform models, explanations can be assigned directly to samples or remapped to time-frequency space. In the representation-comparison study, DFT-LRP inserted a virtual STFT and inverse DFT identity block immediately after the waveform input so that relevance could be visualized in a mel-spectrogram-shaped map without changing the forward computation (Frommholz et al., 2023). In AudioMNIST, waveform explanations were also converted into audible heatmaps by masking the signal with positive relevance, $\tilde x(t)=\mathrm{ReLU}(R(t))\odot x(t)$ (Becker et al., 2018).

For spectrogram models, the explanation target is typically a time-frequency relevance map $R(t,f)$ or an occlusion map over patches. This is the setting used in anomalous sound detection, cough analysis, Grad-CAM-style deepfake explanation, and many prototype methods (Buck et al., 26 Jan 2026, Amado-Caballero et al., 20 Aug 2025, Kitłowski et al., 12 Jun 2026, Kawa et al., 11 May 2026). However, the audio domain also requires more specialized explanation targets. In phoneme recognition, LIME-TS explains the prediction of a particular phoneme instance by ranking fixed-duration time blocks within a sliding window around the phoneme’s ground-truth time (Wu et al., 2023). In transformer-based ASR, LEAF-X assigns each decoded token $y_i$ a normalized attribution vector $s_i\in\Delta^{T-1}$ over acoustic frames, using cross-attention structure, entropy weighting, rollout, and optional causal ablations (Ranjan et al., 12 Jun 2026).

Generative models introduce another target: the contribution of text tokens to generated audio. AudioGenX wraps a frozen text-to-audio model with an Explainer that outputs a soft mask $M_{U,z_t}\in[0,1]^{L\times 1}$ for text-token embeddings at each audio generation step, then aggregates these masks into token-level importance scores across the output sequence (Kang et al., 1 Feb 2025). By contrast, XANE makes the explanation target the latent embedding itself: a 128-dimensional acoustic embedding is explicitly trained to predict 11 regression variables and 3 classification variables, making the embedding explainable in terms of measurable acoustic parameters rather than post-hoc saliency alone (Dumpala et al., 2024).

3. Methodological families

A large part of Audio-XAI consists of adapted vision methods. LRP, Grad-CAM, Integrated Gradients, SmoothGrad, SHAP variants, and LIME appear repeatedly across tasks (Becker et al., 2018, Buck et al., 26 Jan 2026, Grinberg et al., 23 Jan 2025, Wu et al., 2023). Their adaptation is rarely trivial, because audio signals have strong temporal structure, phase-sensitive waveform structure, and acoustically meaningful decompositions that do not align cleanly with image superpixels or visual attention maps.

Perturbation-based methods form one major family. LIME constructs perturbed samples $x'=x\odot z$ and fits a locally weighted surrogate $g(z)=w_0+\sum_{j=1}^d w_j z_j$ by minimizing a locality-weighted square loss plus a sparsity penalty (Wu et al., 2023). The time-partitioned LIME-TS variant replaces manual segments with uniform temporal blocks and a sliding-window perturbation regime centered on the phoneme of interest (Wu et al., 2023). Occlusion methods instead zero out local spectrogram patches or frequency bands and observe output changes; in cough analysis the occlusion map is defined as $M(f,t)=P(y|S)-P(y|S_{\mathrm{occl}}(f,t))$, whereas in anomalous sound detection frequency-band removal produces a functional ground truth for band importance through $$\Delta \mathrm{Pred}_b=|\mathrm{Pred}_{\mathrm{orig}}-\mathrm{Pred}_b|$$ (Amado-Caballero et al., 20 Aug 2025, Buck et al., 26 Jan 2026).

Gradient- and relevance-based methods form a second family. Audio deepfake detection provides a particularly explicit example. Because Wav2Vec2-AASIST has no dedicated classification token, Grinberg et al. adapt transformer relevancy into Gradient Average Transformer Relevancy (GATR), with layer update

$$R_{\text{upd}}=R_{\text{old}}+\bar A_iR_{\text{old}},\qquad \bar A_i=\mathbb{E}_h[(\nabla A_i\odot A_i)^+],$$

followed by a gradient-weighted averaging across token rows and interpolation back to waveform length (Grinberg et al., 23 Jan 2025). This is an audio-specific solution to a transformer architecture that does not fit standard vision-transformer relevance extraction. LEAF-X is similarly architecture-aware: it computes head entropy

$R(t,f)$0

and converts it into a confidence weight $R(t,f)$1 to privilege low-entropy, high-impact heads during token-to-frame attribution (Ranjan et al., 12 Jun 2026).

Prototype and disentanglement methods constitute a third family. APEX inserts an invertible matrix $R(t,f)$2 between a frozen feature map and global average pooling, reparameterizing the classification head as $R(t,f)$3 so that the logits remain identical. It then maximizes channel purity and extracts four types of prototypes: square-based, time-based, frequency-based, and time-frequency hybrid (Kawa et al., 11 May 2026). AudioGenX is conceptually related in that it uses a dedicated Explainer module rather than raw gradients, but its objective is causal text-token selection through factual and counterfactual losses at each audio token (Kang et al., 1 Feb 2025). XANE, finally, is closer to interpretable representation learning than to attribution: its explainability derives from multitask prediction of physically meaningful acoustic parameters (Dumpala et al., 2024).

4. Evaluation protocols, faithfulness criteria, and human interpretability

A central methodological issue in Audio-XAI is that visual plausibility is insufficient. Several works therefore define task-specific quantitative evaluation protocols. In phoneme recognition, explanation quality is measured against TIMIT alignments by asking whether the true phoneme segment appears in the top-$R(t,f)$4 ranked segments, yielding $R(t,f)$5, $R(t,f)$6, and $R(t,f)$7 (Wu et al., 2023). In machine anomalous sound detection, faithfulness is measured by the Spearman correlation between mean band-level relevance $R(t,f)$8 and frequency-band sensitivity $R(t,f)$9 obtained by systematic band removal (Buck et al., 26 Jan 2026). In ASR, LEAF-X uses D-AOPC, insertion, temporal localization, sparsity, stability, and infidelity, combining perturbation sensitivity with forced-alignment grounding and robustness to mild label-preserving perturbations (Ranjan et al., 12 Jun 2026).

In audio deepfake detection, evaluation is especially elaborate. Grinberg et al. define Average Drop, Average Increase, Average Gain, and Input Fidelity on peak-normalized relevance-masked waveforms, then supplement these with positive and negative perturbation tests based on progressively noise-masking the top or bottom $y_i$0 of time frames and measuring the area under the resulting EER curve (Grinberg et al., 23 Jan 2025). They also introduce a partial-spoof localization protocol with Relative Contribution Quantification (RCQ), Relevance Mass Accuracy (RMA), and Relevance Rank Accuracy (RRA) on utterances containing contiguous spoof segments (Grinberg et al., 23 Jan 2025). For text-to-audio generation, AudioGenX defines factual and counterfactual fidelity drops, KL divergences under a frozen audio classifier, and a mask-size measure to capture simplicity (Kang et al., 1 Feb 2025).

Human interpretability is evaluated more rarely, but AudioMNIST provides a notable example. Its user study compared waveform-only, waveform-plus-LRP, raw audio, and masked-audio explanations using informedness and markedness. Audible explanations outperformed visual explanations for both correct and incorrect model outputs, and for incorrectly classified samples the reported informedness and markedness were approximately $y_i$1 and $y_i$2 for audible explanations versus approximately $y_i$3 and $y_i$4 for visual explanations (Becker et al., 2018). A plausible implication is that audio-specific presentation formats may matter as much as the attribution algorithm itself when explanations are intended for human audit.

5. Empirical findings across tasks

The empirical record shows that explanation results are highly task- and representation-dependent. In AudioMNIST, LRP on a spectrogram model for speaker-sex classification concentrated on low frequencies, leading to the hypothesis that the model used differences in fundamental frequency $y_i$5; when male spectrograms were frequency-scaled by $y_i$6 and female spectrograms by $y_i$7, accuracy on the manipulated test set dropped to $y_i$8 (Becker et al., 2018). For the waveform model, zeroing only $y_i$9 of samples selected by LRP relevance reduced digit-task accuracy from approximately $s_i\in\Delta^{T-1}$0 to approximately $s_i\in\Delta^{T-1}$1, larger than amplitude-based or random deletion (Becker et al., 2018).

In audio event classification, relevance maps exposed major differences between raw-waveform and log-mel models trained on UrbanSound8k. The 1D-CNN showed within-class cosine similarity $s_i\in\Delta^{T-1}$2 and between-class similarity $s_i\in\Delta^{T-1}$3, whereas YAMNet showed within-class similarity $s_i\in\Delta^{T-1}$4 and between-class similarity $s_i\in\Delta^{T-1}$5 (Frommholz et al., 2023). The waveform model was also less sensitive to high-pass, low-pass, and pitch-shift perturbations than YAMNet (Frommholz et al., 2023). This suggests that the choice of input representation can alter not only accuracy but also the semantic distinctiveness and robustness of the learned decision strategy.

ASR studies make two complementary points. On TIMIT phoneme recognition, LIME-TS placed the ground-truth phoneme segment in its top three audio segments $s_i\in\Delta^{T-1}$6 of the time and reached $s_i\in\Delta^{T-1}$7 (Wu et al., 2023). On large transformer ASR models, LEAF-X reported D-AOPC $s_i\in\Delta^{T-1}$8 versus a best baseline of $s_i\in\Delta^{T-1}$9, SPR $M_{U,z_t}\in[0,1]^{L\times 1}$0 versus $M_{U,z_t}\in[0,1]^{L\times 1}$1, STAB $M_{U,z_t}\in[0,1]^{L\times 1}$2 versus $M_{U,z_t}\in[0,1]^{L\times 1}$3, and INF $M_{U,z_t}\in[0,1]^{L\times 1}$4 versus $M_{U,z_t}\in[0,1]^{L\times 1}$5, with results summarized as roughly $M_{U,z_t}\in[0,1]^{L\times 1}$6 improved faithfulness, $M_{U,z_t}\in[0,1]^{L\times 1}$7-$M_{U,z_t}\in[0,1]^{L\times 1}$8 stronger locality or sparsity, and the most stable attributions (Ranjan et al., 12 Jun 2026).

Audio deepfake detection reveals both performance differences among explainers and instability of small-sample narratives. On ASV19, GATR achieved fidelity $M_{U,z_t}\in[0,1]^{L\times 1}$9 and Average Drop $x'=x\odot z$0, and on PartialSpoof it reached RMA $x'=x\odot z$1 and RRA $x'=x\odot z$2, outperforming the compared XAI methods (Grinberg et al., 23 Jan 2025). Large-scale RCQ analysis further found that non-speech regions were most influential for bona-fide classification, that low-energy speech frames often carried higher importance except on ITW spoof utterances where high-energy frames dominated, and that unstressed vowels had the highest normalized RCQ on ASV19 spoof files while consonants were least important on ITW (Grinberg et al., 23 Jan 2025). The authors explicitly note that conclusions drawn from a handful of hand-picked samples may not generalize across datasets or conditions (Grinberg et al., 23 Jan 2025).

In anomalous sound detection, low-frequency bands dominated the anomaly model’s decisions: removing Band 1 $x'=x\odot z$3-$x'=x\odot z$4 caused the largest performance drop, whereas removing Band 5 $x'=x\odot z$5-$x'=x\odot z$6 slightly improved Dev AUC to $x'=x\odot z$7 (Buck et al., 26 Jan 2026). Faithfulness evaluation showed Occlusion with $x'=x\odot z$8, higher than Integrated Gradients $x'=x\odot z$9, Grad-CAM $g(z)=w_0+\sum_{j=1}^d w_j z_j$0, and SmoothGrad $g(z)=w_0+\sum_{j=1}^d w_j z_j$1 (Buck et al., 26 Jan 2026). In cough analysis, occlusion maps highlighted mid-frequency bands of approximately $g(z)=w_0+\sum_{j=1}^d w_j z_j$2-$g(z)=w_0+\sum_{j=1}^d w_j z_j$3 around the central portion of the cough epoch, and XAI-weighted spectral features revealed significant group differences for COPD comparisons that were absent in raw spectrogram features (Amado-Caballero et al., 20 Aug 2025). In text-to-audio generation, AudioGenX concentrated importance on sound-related nouns, down-weighted absent optional concepts in some generations, and revealed very low scores for negation tokens in prompts such as “without thunder” and “without no thunder,” exposing a data-bias in the TAG model (Kang et al., 1 Feb 2025).

6. Fragility, interpretable-by-design alternatives, and open problems

A recurring concern is whether post-hoc audio explanations are stable under innocuous input changes. "The Perceived Fragility of Explanations in Audio Models" formalizes this concern by optimizing an inaudible perturbation $g(z)=w_0+\sum_{j=1}^d w_j z_j$4 that minimizes explanation cosine similarity while preserving the original prediction under a hinge-loss constraint (Kitłowski et al., 12 Jun 2026). The reported attacks retained high perceptual quality, preserved the final “real” or “fake” label with $g(z)=w_0+\sum_{j=1}^d w_j z_j$5 preservation under the hinge loss, reduced explanation cosine similarity from approximately $g(z)=w_0+\sum_{j=1}^d w_j z_j$6 to approximately $g(z)=w_0+\sum_{j=1}^d w_j z_j$7, and yielded median $g(z)=w_0+\sum_{j=1}^d w_j z_j$8 across VGGish, AST, and SpecTTTra (Kitłowski et al., 12 Jun 2026). The paper also recommends explanation adversarial training, multiple complementary XAI methods, perceptual consistency constraints, decision-tethered explanations, and deployment-time monitoring of fragility (Kitłowski et al., 12 Jun 2026).

One response to such fragility is to make explanations more tightly coupled to model structure. LEAF-X is model-intrinsic for transformer ASR and uses entropy-guided attention weighting plus optional causal layer ablations (Ranjan et al., 12 Jun 2026). APEX preserves logits exactly through an invertible reparameterization and offers prototype explanations from four audio-centric perspectives without fine-tuning the original backbone (Kawa et al., 11 May 2026). XANE goes further toward explainability-by-design by forcing a compact embedding to reproduce 14 acoustic parameters and reporting a mean F1 score of $g(z)=w_0+\sum_{j=1}^d w_j z_j$9 on three clustering tasks, alongside parameter estimation and a CPU real-time factor of $M(f,t)=P(y|S)-P(y|S_{\mathrm{occl}}(f,t))$0 (Dumpala et al., 2024). These approaches do not eliminate the need for evaluation, but they reduce the conceptual gap between internal computation and explanation.

Several limitations recur across the literature. APEX requires a pooling-to-linear classifier head and does not yet handle transformer backbones or self-supervised representations directly (Kawa et al., 11 May 2026). LEAF-X experiments are confined to English-only Whisper and Canary models (Ranjan et al., 12 Jun 2026). AudioGenX currently does not address diffusion-based text-to-audio models without explicit discrete tokens (Kang et al., 1 Feb 2025). AudioMNIST notes that spectrogram explanations were not made audible because waveform reconstruction would require phase reconstruction (Becker et al., 2018). The cumulative evidence suggests that Audio-XAI remains heterogeneous: explanation faithfulness depends on task, architecture, and representation, and explanation usability depends on both quantitative validation and audio-appropriate presentation.