Study on elderly speech recognition for achieving speech interfaces in ubiquitous computing environment

语音识别技术作为老年人友好的人机界面之一备受关注。然而，与非老年成人语音相比，老年语音会导致语音识别性能的大幅下降。本研究以提高老年语音的识别率为目的，通过声学分析对老年语音的声学特征进行了定量的检测，重点研究了老年语音的“非轻快”和“沙哑”特征，成功地找到了听力测试给出的主观特征与声学分析得到的客观特征之间的关系。对于非轻快的老年语音，主观非轻快是由于年龄的原因导致发音器官的模糊运动所致。在这项研究中，我们发现主观非快感的程度与后续音素之间频谱包络的时间移动有关。此外，语音功率的时间演化与主观非轻快以及频谱包络的时间运动有关，对于沙哑的老年嗓音，我们认为老年声带上的噪声给人留下了主观沙哑的印象。我们比较了老年人和非老年人发出的正常嗓音对每个元音的平均幅度谱，发现老年人嘶哑的声音在中频(1.5 kHz-2.5 kHz)和2.5 kHz以上的高频范围分别有功率降低和功率增加的现象。我们用振幅谱的倾斜度对这一现象作了简要的解释。使用预处理器进行语音识别，将老年语音的频谱倾斜度调整为非老年成人的频谱倾斜度，并验证了预处理器在元音识别任务中的良好工作。因此，我们得出结论，声音沙哑是老年人语音识别率较低的原因之一。

Speech recognition technology is attractive as one of friendly human-machine interfaces for elderly people. However, elderly speech causes a great decrease in the performance of speech recognition compared to non-elderly adult speech. This study aims at improving the recognition rate of elderly speech, and carries out acoustic analysis to examine the acoustic characteristic of elderly speech quantitatively.This study focused on "non-briskness" and "hoarseness" as the nature of elderly speech, and successfully found the relationship between subjective characteristics given by listening tests and objective features obtained by acoustic analysis.Concerning the non-brisk elderly voice, it is well known that subjective non-briskness is caused by the vague movements of the articulatory organs due to aging. In this study, we found that the degree of subjective non-briskness related to the temporal movement of spectral envelops between succeeding phonemes. Furthermore, time evolution of speech power has the relationship with the subjective non-briskness as well as the temporal movement of spectral envelopes.Concerning the hoarse elderly voice, we consider that noise occurred at the aged vocal cords impresses us the subjective hoarseness. We compared the averaged amplitude spectra between elderly hoarse voices and normal voices uttered by non-elderly adults for each of Japanese vowels, and found that elderly hoarse voice had power decrease and increase in the mid frequency range (1.5 kHz-2.5 kHz) and the high frequency range over 2.5 kHz, respectively. We explain this phenomenon briefly by using tilt of amplitude spectrum. Speech recognition was carried out with the preprocessor, which adjusts the spectral tilt of elderly voice to that of non-elderly adult, and we confirmed the preprocessor worked well on the vowel recognition task. Therefore, we conclude that the hoarseness is one of the reasons for the worse recognition rate of elderly speech.