Speech basis fashions have not too long ago achieved exceptional capabilities throughout a variety of duties. Nonetheless, their analysis stays disjointed throughout duties and mannequin varieties. Totally different fashions excel at distinct elements of speech processing and thus require completely different analysis protocols. This paper proposes a unified taxonomy that addresses the query: Which analysis is suitable for which mannequin? The taxonomy defines three orthogonal axes: the analysis side being measured, the mannequin capabilities required to try the duty, and the duty or protocol necessities wanted to carry out it. We classify a broad set of present evaluations and benchmarks alongside these axes, spanning areas reminiscent of illustration studying, speech era, and interactive dialogue. By mapping every analysis to the capabilities a mannequin exposes (e.g., speech era, real-time processing) and to its methodological calls for (e.g., fine-tuning information, human judgment), the taxonomy supplies a principled framework for aligning fashions with appropriate analysis strategies. It additionally reveals systematic gaps, reminiscent of restricted protection of prosody, interplay, or reasoning, that spotlight priorities for future benchmark design. Total, this work presents a conceptual basis and sensible information for choosing, deciphering, and lengthening evaluations of speech fashions.
