Acoustic correlates of phonetic perception

描述（由申请人提供）：如果我们对语音识别背后的听觉和感知机制有更深入的了解，许多实际和理论问题可以得到解决。这些知识的实际应用包括人工耳蜗信号处理器的改进、语音合成设备的改进、鲁棒语音识别算法的开发以及听力受损说话者训练设备的开发。提出的实验分为三大类。一系列广泛的实验旨在测试元音感知新模型的有效性，该模型是在上一个资助期间开发的。该模型假设元音身份是通过模板匹配过程来识别的，该过程涉及将窄带输入光谱与一组平滑的光谱形状模板进行比较，这些模板是通过普通的语音暴露来学习的。在上一个拨款期间进行的评估表明，该模型能够从一个大型的多语数据库中识别元音，其准确性接近人类听众。我们希望扩大这一工作范围，以解决以下问题：(1)建立时间和频谱线索对元音同一性的整合模型；(2)测试该理论对语音环境、信号周期性和频谱形状特征（如峰振幅关系）等因素变化的鲁棒性；(3)结合一种心理上合理的归一化方案，该方案可能解释听者识别具有不同声道特征的说话者的能力。第二组实验将进行，测量有多少语音信息通过语音信号传递给听众，这些语音信号是由专门设计的语音合成算法产生的，该算法只保留原始语音信号的某些特征，而有意去除或扭曲其他线索。结果将允许推断出有关调解语音识别的潜在频谱表示的性质。第三项工作旨在评估听觉频率分析的新理论，该理论是在上一个拨款期间发展起来的。该理论认为，通过对听神经放电模式的分析，听觉频谱不是在外围神经系统计算的，而是在中枢神经系统计算的。一个软件模拟将被开发用于测试理论的有效性。

DESCRIPTION (provided by applicant): Numerous practical and theoretical problems could be addressed if we had a deeper understanding of the auditory and perceptual mechanisms underlying phonetic recognition. Practical applications of this knowledge include improvements to cochlear implant signal processors, the improvement of speech synthesis devices, the development of robust speech recognition algorithms, and the development of training devices for hearing-impaired speakers. The proposed experiments fall into three major categories. An extensive series of experiments is designed to test the validity of a new model of vowel perception that was developed during the previous grant period. The model assumes that vowel identity is recognized by a template-matching process involving the comparison of narrow band input spectra with a set of smoothed spectral-shape templates that are learned through ordinary exposure to speech. An evaluation conducted during the previous grant period showed that the model is capable of recognizing vowels from a large, multitalker database with accuracy approaching that of human listeners. We would like to extend this line of work to address issues such as: (1) modeling the integration of temporal and spectral cues to vowel identity, (2) testing the robustness of the theory to variation in factors such as phonetic environment, signal periodicity, and spectral shape features such as formant amplitude relations, and (3) incorporation of a psychologically plausible normalization scheme that might account for the ability of listeners to recognize speech from talkers with diverse vocal-tract characteristics. A second set of experiments will be run which measure how much phonetic information is transmitted to listeners by speech signals that have been generated by specially designed speech synthesis algorithms that preserve only some characteristics of the original speech signal, while purposely removing or distorting other cues. The results will allow inferences to be drawn about the nature of the underlying spectral representations that mediate phonetic recognition. A third line of work is designed to evaluate a new theory of auditory frequency analysis that was developed during the previous grant period. The theory assumes that the auditory spectrum is calculated not at the periphery but in the central nervous system through an analysis of auditory nerve firing patterns. A software simulation will be developed for use in testing the validity of the theory.