计算听觉场景分析及计算机听觉模型的研究

This project presents a concurrent vowels separation system.based on double-pitch multi-band excitation separation model by using.auditory periphery model including outer and middle ear resonances,.cochlear filtering and neuromechanical transduction. A new method to.provide even more effective reference inputs for adaptive signal enhancer.is presented in the chinchilla model, sepecifically, a Gaussian radial.basis function neural network was used to preprocess raw evoked potentials.signals. The effects on hearing and the sensory cell population of four.continuous, non-Gaussian noise exposures were compared to the effects of.an energy-equivalent Gaussian noise. An energy metric in combination with.the statistical metric, kurtosis,was presented to provide necessary.information on any noise environment to evaluate its potential for causing.hearing loss and cochlear sensory cell loss. Correlations among evoked.potential thresholds, distortion product otoacoustic emissions and hair.cell loss following various noise exposures in the chinchina was studied.

本项目提出了主动型ＣＡＳＡ系统的研究方向，利用声音的固有属性将声音分为三个基本模型，对输入的混合声进行预测、修正、分析归类；利用声音的空间属性研究基于ＨＲＴＦ的声源空间定位，并将它作为声源分离的一个重要线索；试图利用ＣＡＳＡ解决语音识别系统在复杂环境下应用的难题。本项目对计算机听觉等研究有重要意义，并有广阔的应用前景。