Based on a handful of basic measurements, it has been shown possible to confidently predict the subjective judgement of these loudspeakers. New input features have been developed and combined with existing sound metrics and raw measurements to provide the inputs. The program utilises cascaded ensemble methods, and is simplified to ensure a level of complexity appropriate to the data. Automatic anomaly detection removes potentially misleading data before recalculating the most suitable decision boundary, based on those loudspeakers most important to the groupings. The final performance based on separate test data shows a marked improvement relative to suitable alternative methods, and a drastic improvement compared to analysis of any one measurement alone. The output is a classification which correlates highly to subjective judgements.
展开▼