Variable-size Vector Entropy Coding of Speech and Audio

Source coding is an essential operation in modern digital communication and multimedia networks. Many modern coders employ some form of entropy coding (EC) to lower the transmission rate below the capacity limits of the communication channel. Usually, a simple form of entropy coding is used in order to meet the constraints on the complexity and price of the encoder and decoder. To maintain low complexity, EC is applied to symbols of small alphabets where the symbols represents a single element (typically, a quantization index) or a very small vector of elements (vector-EC). In this work we extend the idea of vector-EC using vectors of unconstrained sizes which may be as small as 1 and as large as several hundreds. The proposed method is, however, complexity-constrained in the sense that the symbol vector size is always as large as allowed by a complexity limit. The method is studied in the framework of an MDCT transform coder. It is shown experimentally, using diverse audio material, that a coding rate reduction of about 37% can be achieved, relative to the rate of a scalar Huffman coder. The method is, however, not specific to MDCT coding but can be incorporated in various speech audio image and video coders.