SpeechDx Benchmark Advances Clinical Speech AI Evaluation

A new comprehensive benchmark, SpeechDx, has been introduced to standardize and advance the evaluation of clinical speech AI. This benchmark integrates 12 diverse datasets and 27 tasks, covering a wide range of health conditions that affect speech. SpeechDx organizes tasks based on the stage of speech production they disrupt—conceptualization, formulation, and articulation—to facilitate a more structured assessment of AI models. This approach aims to test generalization capabilities, especially with limited labeled data, and to distinguish genuine clinical patterns from dataset-specific anomalies. Systematic evaluation of 12 state-of-the-art audio encoders revealed that large-scale speech models generally perform best, though domain-specific models can excel in closely matched tasks. Critically, no current representation reliably generalizes across the entire clinical speech landscape, underscoring the need for further research into robust, general-purpose clinical speech representations.