Weisenberger J M, Craig J C, Abbott G D
Central Institute for the Deaf, St. Louis, Missouri 63110.
J Acoust Soc Am. 1991 Oct;90(4 Pt 1):1944-57. doi: 10.1121/1.401674.
Principal component analysis, a statistical data reduction technique which can be used to eliminate redundant information, has shown promising results as a speech coding strategy in auditory perceptual studies. The present study describes the development, modification, and evaluation of a principal components-based tactile aid for speech perception by the hearing-impaired. In this device, the first two principal components of an input speech signal were displayed on two-dimensional arrays of vibrators contacting either the fingertip or the forearm. Initial testing of the device with closed-set recorded speech tokens showed fair recognition performance, reaching 57% for three consonants and 56% for four vowels. Modifications to the processor algorithm designed to improve vowel recognizability resulted in higher levels of performance (66% for eight vowels). A real-time prototype was constructed implementing the revised algorithm. Live-voice testing was conducted with six normal-hearing subjects, three of whom had previous training with the Queen's University vocoder, a multichannel tactile vocoder that has shown promising results. Performance of these "trained" subjects for both single-item and connected speech tasks was excellent, equalling levels obtained with the Queen's vocoder. These results suggest that a principal components design may be a promising alternative to a vocoder strategy for a tactile aid. Results for the "naive" subjects did not reach the levels attained by the trained subjects, a finding partially attributed to the short training period available to the naive subjects. The higher level of performance for the trained subjects, together with the similarity of performance for the principal components aid and the Queen's vocoder for these subjects, suggests that they were able to transfer previous learning with the Queen's vocoder to the principal components device.
主成分分析是一种可用于消除冗余信息的统计数据简化技术，在听觉感知研究中作为一种语音编码策略已显示出有前景的结果。本研究描述了一种基于主成分的触觉辅助设备的开发、改进和评估，该设备用于帮助听力受损者进行语音感知。在这个设备中，输入语音信号的前两个主成分显示在与指尖或前臂接触的二维振动器阵列上。使用封闭式录制语音样本对该设备进行的初步测试显示出了不错的识别性能，三个辅音的识别率达到57%，四个元音的识别率达到56%。对旨在提高元音可识别性的处理器算法进行修改后，性能水平更高（八个元音的识别率为66%）。构建了一个实现修订算法的实时原型。对六名听力正常的受试者进行了现场语音测试，其中三名受试者之前接受过女王大学声码器的训练，女王大学声码器是一种多通道触觉声码器，已显示出有前景的结果。这些“受过训练”的受试者在单项和连贯语音任务中的表现都非常出色，与女王大学声码器获得的水平相当。这些结果表明，主成分设计可能是触觉辅助设备中声码器策略的一个有前景的替代方案。“未受过训练”的受试者的结果未达到受过训练的受试者所达到的水平，这一发现部分归因于未受过训练的受试者可用的训练时间较短。受过训练的受试者的较高表现水平，以及主成分辅助设备和女王大学声码器对这些受试者的表现相似性，表明他们能够将之前使用女王大学声码器的学习经验转移到主成分设备上。
