符號冗息在訊源控制通道解碼之研究

Abstract

在數位傳輸系統的設計中，合併訊源通道解碼的研究旨在補償通道錯誤所造成的解碼失真。訊號通過訊源編碼器之後，在量化輸出位元串之間存在某種形式的關聯性，若能在通道解碼器妥善地利用這一項事前資訊，將可對接收端的解碼能力有所提升。我們提出了一個訊源控制通道解碼機制，在接收端將利用量化後所殘留的訊源冗息提昇通道解碼工作的正確性。在這架構中，有別於以往以單一位元為基礎的演算法，我們探討以符號為層級的改良式BCJR演算法，計算每個符號的後驗機率後再做最大後驗機率的估測。有關論文的實驗模擬，先以自迴歸訊號源做驗證，再將此架構應用於現今被廣泛地使用的分散式語音辨識系統上作討論分析。

For digital multimedia communication, the system performance will degrade after compressed signals are transmitted over noisy channels. In order to combat channel errors, the signals will be protected by means of joint source-channel coding methods. When signals are quantized by the source encoder into bit sequences, there exists residual redundancy, due to non-uniform distribution and correlations between them. By taking advantage of this a priori knowledge, we propose a source-controlled channel decoding scheme which incorporates the residual redundancy into the channel decoding process. Also proposed is a modified BCJR algorithm that computes the a posteriori probability at symbol level, as opposed to conventional bit-level BCJR algorithm. Experiments were carried out to investigate the robustness of the proposed coding approach. The input signals considered here include first-order autoregressive sources and speech parameters used for distributed speech recognition.