以基因法則設計通道匹配向量量化之研究

Abstract

向量量化是一項重要的影音資料壓縮技術。然而傳統的碼書設計方式 ，卻會在訓練的過程中陷入區域最佳值。本論文首先希望應用基因法則的 全域平行隨機搜尋特性，利用染色體配對與基因突變的遺傳演化設計更有 效的向量量化碼書。接著我們將基因演化的設計成功地延伸，以有效對抗 通道雜訊的碼書訓練。另外，我們也發現能匹配通道雜訊的碼字索引排列 也能減小碼字傳送的平均失真。針對這個索引配置問題，我們結合了二元 交換技術以及基因演算法則來求取最佳的索引配置，並妥善處理染色體編 碼模式，以避免基因配對與突變所衍生的不合法索引排列。在本論文的最 後，將我們發展的碼書訓練技術，應用在交大語音傳輸實驗室建構的電子 郵件語音轉換系統上。 Vector quantization (VQ) has been widely used in speech and image coding for data compression. It operates by encoding a sequence of input vectors with a codebook and by transmitting the index of the nearest codevector to the receiver.Thus, the effects of channel errors on transmitted codevector indices can result in significant distortion in decoded vectors. In this work we explore the benefits of genetic algorithm for use in designing a VQ with increased robustness to channel errors. The first part of this investigation is concerned with the accuracy of the codebook training. The well-known GLA algorithm iteratively updates the codebook according to the nearest neighbor and centroid conditions. However, its simple downhill search transitions can easily become trapped in local optima and thus miss finding the globally optimal codebook. To compensate this shortage, we develop a codebook training technique based on the genetic algorithm. The main attraction of genetic algorithm arises from the fact that the given search space is explored in parallel by means of iterative modifications of a population of potential solutions (chromosomes). An alternative approach to achieving robustness is that of index assignment. Here the redundancy-free error protection is achieved by using a hybrid strategy which merges the most appealing features of two algorithms, the genetic algorithm helps to identify the hills and the binary switching algorithm goes and climbs them. To match the index assignment problem, the encoding of chromosomes must be carefully designed to eliminate the illegality of the offsprings following the crossover and mutation operators. Finally, we apply the enhanced VQ technique to develop a voiced mail system that uses touch tones as input and synthetic speech as output to provide remote telephone access to electronic mails.