Context:sensitivity/bias/parsing phonetic information

描述：(申请人提供)：在流利的演讲中，演讲者在读完最后一个音之前就开始发下一个音。其结果是，语音不仅一个接一个地出现，而且实际上是重叠的，停顿只发生在整个短语之间，而不是单个声音。流利语音的这一特点给听者带来了两个棘手的问题：分离重叠的声音，然后识别由于与邻居的发音重叠而音质失真的声音。这一建议遵循这样的假设，即分离和识别都可以发生，因为信号中的连续间隔在感知上相互反差。例如，在其中大部分声能在高频的间隔之后，能量在中频的声音将听起来相对较低，或者在相对较长的间隔之后，中间持续时间的间隔将听起来相对较短。实验测试了这种假设的一个版本，其中在听者为声音分配任何语言值之前，即在声音被识别为特定类别的实例之前，在声音的初始听觉评估中像这样夸大顺序对比。如果在识别声音之前出现顺序对比，则它将不受收听者可能具有的任何语言知识的影响，例如当前声音是否与其上下文一起构成单词、在该上下文中频繁出现、在该上下文中在语音上合法等。通过更好地区分在其顺序对比方向上不同的声音序列，例如，高-低和低-高，而不是不同的序列，即高-高和低-低，来诊断语音处理的单独的、前语言的、听觉的阶段。如果在加工的所有阶段都使用语言知识，这两对序列应该同样容易区分，因为所有的间隔都将被分配到类别，因此将同样不同。因此，这些实验的结果允许在语音识别的交互模型和自主模型之间进行选择，在该交互模型中，收听者在处理他们听到的语音时在所有阶段使用他们的语言知识，在该自主模型中，他们在第一阶段期间仅使用信号的心理声学属性，并且仅在稍后将他们在语言上所知道的应用于该阶段的输出。如果支持自主模型，那么通过提高信号质量而不是增加冗余的语言信息，可以更好地提高在不利条件下或受损听者的语音感知的稳健性。

DESCRIPTION: (provided by applicant): In fluent speech, speakers begin to pronounce the next sound before they're done pronouncing the last. As a result, speech sounds not only occur right next to one another but actually overlap, and pauses only occur between whole phrases and not individual sounds. This characteristic of fluent speech presents the listener with two formidable problems: separating overlapping sounds and then recognizing sounds whose acoustics have been distorted by the overlap with its neighbors' pronunciations. This proposal pursues the hypothesis that both separation and recognition can happen because successive intervals in the signal contrast with one another perceptually. For example, after an interval in which most of the sound energy is at high frequencies, a sound whose energy is at mid frequencies will sound relatively low, or after a relatively long interval, an interval of intermediate duration will sound relatively short. The experiments test a version of this hypothesis in which sequential contrast is exaggerated like this in the initial auditory evaluation of the sounds, before the listener has assigned any linguistic value to the sound, i.e. before the sounds are recognized as instances of particular categories. If sequential contrast arises before the sounds are recognized, then it will be impervious to any linguistic knowledge the listener may have, e.g. of whether the current sound makes a word with its context, occurs frequently in that context, is phonotactically legal in that context, etc. A separate, prelinguistic, auditory stage of phonetic processing is diagnosed by better discrimination of sound sequences that differ in the direction of their sequential contrast, e.g. high-low vs low-high, than of sequences that don't, i.e. high-high vs low-low. If linguistic knowledge is used at all stages of processing, these two pairs of sequences should instead be equally easy to distinguish because all the intervals will have been assigned to categories and will therefore be equally different. The results of these experiments therefore permit a choice between interactive models of speech sound recognition in which listeners use their linguistic knowledge at all stages in processing the speech sounds they hear and autonomous models in which they use only the psychoacoustic properties of the signal during the first stage, and only later apply what they know linguistically to the output of that stage. If the autonomous model is supported, then the robustness of speech perception under adverse conditions or by impaired listeners can be improved more by enhancing signal quality than adding redundant linguistic information.