Audio transcoding is the conversion of digital audio from one compressed form A to another compressed form B, where A and B have different compression properties, such as a different bit-rate, sampling frequency or compression method. This is typically achieved by decoding A to an intermediate uncompressed form, and then encoding it to B. A significant portion of the involved computational effort pertains to operating the synthesis filter bank, which is an important processing block in the decoding stage, and the analysis filter bank, which is an important processing block in the encoding stage. This thesis presents methods for efficient implementations of filter banks and audio transcoders, and is separated into two main parts. In the first part, a new class of Frequency Response Masking (FRM) filter banks is introduced. These filter banks are usually characterized by comprising a tree-structured cascade of subfilters, which have small individual filter lengths. Methods of complexity reduction are proposed for the scenarios when the filter banks are operated in single-rate mode, and when they are operated in multirate mode; and for the scenarios when the input signal is real-valued, and when it is complex-valued. An efficient variable bandwidth FRM filter bank is designed by using signed-powers-of-two reduction of its subfilter coefficients. Our design has a complexity an order lower than that of an octave filter bank with the same specifications. In the second part, the audio transcoding process is analyzed. Audio transcoding is modeled as a cascaded quantization process, and the cascaded quantization of an input signal is analyzed under different conditions, for the MPEG 1 Layer 2 and MP3 compression methods. One condition is the input-to-output delay of the transcoder, which is known to have an impact on the audio quality of the transcoded material. Methods to reduce the error in a cascaded quantization process are also proposed. An ultra-fast MP3 transcoder that requires only integer operations is proposed and implemented in software. Our implementation shows an improvement by a factor of 5 to 16 over other best known transcoders in terms of execution speed.
|Qualification||Doctor of Philosophy|
|Award date||19 Mar 2007|
|Place of Publication||Eindhoven|
|Publication status||Published - 2007|