The NIST 1999 Speaker Recognition Evaluation An Overview

摘要: Martin, Alvin, and Przybocki, Mark, The NIST 1999 Speaker Recognition Evaluation?An Overview, Digital Signal Processing10(2000), 1?18.This article summarizes the Evaluation. It discusses overall research objectives, three task definitions, development evaluation data sets, specified performance measures their manner of presentation, quality results. More than a dozen sites from United States, Europe, Asia participated in this evaluation. There were primary tasks for which automatic systems could be designed: one-speaker detection, two-speaker speaker tracking. All performed context mu-law encoded conversational telephone speech. detection used single channel data, while other two summed two-channel data. About 500 target speakers specified, with 2 min training speech provided each. Both multiple test segments selected about 2000 conversations that not material. duration was nominally 1 min, varied near zero up to 60 s. For each task, had make independent decisions combinations segment hypothesized speaker. sets designed large enough provide statistically meaningful results on subsets interest. Results analyzed respect various conditions including duration, pitch differences, handset types.