基於聽覺感知模型的混響時間盲估計技術

Abstract

混響時間 (reverberation time, RT) 是描述房間特性中極為重要且廣為使用的聲 學參數之一。它不僅可以提供拿來做為評斷室內混響嚴重程度的衡量，更可以利用 來消除混響，因此，如何準確地估計混響時間是相當重要的議題之一。本論文中， 我們使用一已被提出的，模擬聲音沿著人耳到大腦傳輸路徑的聽覺感知模型，藉由 此模型大腦皮質階段的分析，觀察出其不同混響時間下語音的分佈趨勢。基於這個 現象，以此模型中的大腦皮質階段來抽取語音特徵參數，並透過高斯混合模型 (GMM) 來訓練出不同混響時間的高斯混合模型，再對混響語音進行辨識測試，並利 用不同類別模型之間輸出分數的比例，進而達到連續性的混響時間估計，最後更進 一步延伸到後端混響消除之應用，並證實其估計方法一定的準確性。

The reverberation time (RT) is one of the most prominent acoustic features of an enclosure. It can be used to measure the degree of reverberation and used by some speech enhancement techniques to suppress reverberation. Therefore, it is important to estimate RT accurately. In this thesis, we investigate an auditory perceptual model in the RT estimation application. The auditory model simulates signal processing principles of human hearing along the auditory pathway from the ear to the brain. We observe that the rate-scale plot produced by the second stage of the auditory model encodes the degree of reverberation. Based on this key observation, speech perceptual features extracted from the second stage are used to estimate RT. A perceptual-feature based discrete RT recognizer is built using the Gaussian mixture model (GMM). We then extend the discrete outcomes of the recognizer to produce a continuous value of RT using a probability weighting function. In the end, we develop a method to estimate RT blindly. Experimental results demonstrate the performance of the proposed method.