PHYSIOLOGICAL AND TECHNOLOGICAL STUDIES ON THE AUDITORY SCENE ANALYSYS

即使在噪音掩盖语音的嘈杂环境中，人类也能识别语音。这种现象被称为“鸡尾酒会效应”，或者Bregman称之为“听觉场景分析”（1991）。他在他的心理物理学研究中提出，这种效应背后有几种听觉机制，如声音的时间感应（在被噪音掩盖的时期恢复原始声音）、声音的分割、分组和定位。这些机制在生理学上还没有被完全理解。本研究旨在探讨听觉皮层听觉场景分析的生理机制，并建立听觉皮层听觉场景分析的神经模型。利用神经活动的光学成像技术研究了豚鼠听觉皮层(1)时间诱导、(2)分割分组和(3)声音定位的神经机制；(4)听觉皮层神经模型和噪声鲁棒性语音识别模型的研究。对插入间隙的纯音的反应，在间隙前后音调开始时出现更多的红色。插入噪声抑制了对间隙后音调的起始反应，并且这种抑制在带场中比在人工智能中更强。这些结果表明，抑制噪声后的起始检测是听觉诱导的重要机制。在高调频下，人工脑和前带场的响应服从于调幅声（AM），而在低调频下，背尾带场和尾带场的响应服从于调幅声（AM），且高调频下尾带场的偏移响应更大。这些结果表明，尾带区域处理声音的时间窗口较长，表明该区域对复杂声音的分析更加专业化，并表明偏移响应参与了声音分组机制。与人工智能相比，尾带场对声音位置更敏感。听觉皮层的神经模型再现了人工智能的时空神经活动。少

Human can recognize speeches even in noisy environments where noise masks speech sounds. This phenomenon is called 'cocktail party effect' or, by Bregman, 'auditory scene analysis' (1991). He suggested in his psychophysical studies that there are several auditory mechanisms underlying this effect such as temporal induction of sounds (restoration of original sounds in the periods masked by noise), segmentation, grouping and location of sounds. These mechanisms are not fully understood physiologically. The present study is aimed to investigate physiological mechanisms of auditory scene analysis in the auditory cortex and to make a neural model. Neural mechanisms of (1) temporal induction, (2) segmentation and grouping and (3) sound location were investigated in the guinea-pig auditory cortex using an optical imaging technique of neural activity and (4) a neural model of the auditory cortex and a noise-robust speech recognition model were studied. Responses to gap-inserted pure tone appea … More red at the onset of pre- and post-gap tone. The onset response to the post-gap tone was inhibited by insertion of noise, and this inhibition was stronger in the belt fields than in the AI. These results suggest that the inhibition of the onset detection after the noise is an important mechanism for auditory induction. The response in AI and the anterior belt field followed to the amplitude-modulated sounds (AM) at high modulation frequencies (fm) but the dorsocaudal field and the caudal belt field followed to AM at low fm, and the caudal belt field elicited the larger offset responses for the higher fm. These results show that the caudal belt fields have a longer time window to process sounds, suggesting that this field is more specialized to analyze complex sounds, and show that the offset response is involved in the sound-grouping mechanisms. The caudal belt fields were found to be more sensitive to the sound location than the AI. A neural model of the auditory cortex reproduced spatiotemporal neural activity of the AI. Less