Using written language to probe speech recognition models

语音识别模型分为两大类，它们的处理体系结构根本不同。反馈模型（如TRACE， McClelland & Elman, 1986）允许词汇知识对音素分析施加自上而下的控制。前馈模型（如Merge， Norris, McQueen & Cutler, 2000）假设信息流完全是自下而上的。我们的项目采用了一种创新的方法来测试这些模型类别，通过检查书面单词知识对语音感知的影响。为了区分模型，对比必须检验不同的加工水平和检验策略效应。TRACE倾向于在有限的战略影响下产生广泛的影响；合并有利于对策略因素敏感的词汇效应

Speech recognition models fall into two principal classes, with fundamentally different processing architectures. Feedback models (e.g. TRACE, McClelland & Elman, 1986) allow lexical knowledge to exert top-down control over phonemic analysis. Feedforward models (e.g. Merge, Norris, McQueen & Cutler, 2000) assume that information flow is entirely bottom-up. Our project adopts an innovative approach to testing between these model classes, by examining the influence of written-word knowledge on speech perception. To distinguish the models, contrasts must test different processing levels and examine strategy effects. TRACE favors broad effects with limited strategic influence; Merge favors lexical effects that are necessarily sensitive to strategic factors