Introduction

High quality CD or DAT audio require large amounts of data. A CD stream consists of 44100 16-bit samples per channel ans second, which corresponds to 1.4 Mbit/s in stereo. Audio at that bitrate contains a lot of redundancies, which can be exploited in a lossless coder. A lossless coder does not change the data, but the original audio is completely recovered after decompression. Lossless audio coders can get the bitrate down to about half the original rate, or 700 kbit/s. The human auditory system, though, has many limitations - we simply do not hear a lot of the information in a high quality audio signal. Thus, audio data has a lot more perceptual redundancies, and a lossy coder which exploits those can be made much more efficient. 10 to 12 times less bits than the original can often be used without perceptual loss.

In this project, I have done some research in how perceptual audio coding works, and what is currently done in the field. With concepts from studied papers and standards (and some new ideas), I have implemented an experimental audio transform coder. The coder is not intended to be state-of-the-art, but rather a tool for me to learn the difficulties that arise in a coder of this kind. The report is structured in the following way:

• 2 Existing Coders and Standards describes some of the well-known implementations and standards that exist, and some ideas from those.
• 3 Human Audio Perception: Masking describes the masking properties of the human auditory system, and the implemented model of this in the coder.
• 4 Audio Coding goes through and motivates the quantization and bit coding used in the coder.
• 5 Results and Conclusions shows bitrates and ``quality'' of some encoded audio clips. The audio clips are presented on the web.