Interpretability Audits of Speech-Based Triage Classifiers Under Code-Switching Induced Tokenization Drift: A Computer Robustness Evaluation

This is the authors' abstract. We don't add key points for this paper.

Speech-based triage classifiers that operate on code-switched clinical speech face a structural threat that standard interpretability audits do not detect: the tokenizer that converts audio-derived transcripts into model inputs can shift its segmentation behaviour when speakers alternate between languages, and the resulting tokenization drift silently relocates the very features that attribution methods are supposed to explain. the present analysis formalises that threat and evaluates it as a design problem rather than an empirical claim about deployed systems. It defines tokenization drift as a divergence between the token sequence a fixed tokenizer produces on monolingual and code-switched renderings of the same clinical utterance, and it defines interpretability-audit validity as the stability of feature-attribution rankings under that divergence. A before-after-control-impact design is then characterised analytically under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The design yields 52 clusters for 5 percentage points at 80% power. The contribution is a formal account of why attribution-stability audits of speech triage classifiers are confounded by tokenizer behaviour, together with a design template that separates tokenizer-induced variance from classifier-induced variance before any audit result is interpreted.