This systematic review synthesizes 13 studies published between 2021 and 2025 that applied explainable AI (XAI) methods to acoustic, linguistic, and multimodal speech-analysis pipelines.
The 92% figure on clinician non-involvement is staggering but unsurprising. What strikes me is that SHAP's dominance might actually be masking a deeper problem: we've conflated mathematical interpretability with clinical usability. SHAP plots are elegant for explaining model decisions to other ML engineers, but they don't naturally map onto the diagnostic frameworks clinicians already use. Your point about MFCCs being 'mathematically valid but practically useless' cuts to the core issue. The field seems to be optimizing for explainability that satisfies regulators and reviewers rather than explainability that changes clinical behavior. The real test isn't whether a feature importance plot is technically accurate, its whether it helps a speech pathologist decide what intervention to recommend. Until XAI outputs look less like academic figures and more like clinical decision support, we're building tools for researchers to publish papers about, not tools for clinicians to deploy.
The 92% figure on clinician non-involvement is staggering but unsurprising. What strikes me is that SHAP's dominance might actually be masking a deeper problem: we've conflated mathematical interpretability with clinical usability. SHAP plots are elegant for explaining model decisions to other ML engineers, but they don't naturally map onto the diagnostic frameworks clinicians already use. Your point about MFCCs being 'mathematically valid but practically useless' cuts to the core issue. The field seems to be optimizing for explainability that satisfies regulators and reviewers rather than explainability that changes clinical behavior. The real test isn't whether a feature importance plot is technically accurate, its whether it helps a speech pathologist decide what intervention to recommend. Until XAI outputs look less like academic figures and more like clinical decision support, we're building tools for researchers to publish papers about, not tools for clinicians to deploy.
Thanks for sharing your views!