Audio Coding

Abstract

Speech and audio signals can be characterized as waveforms, which often include components of quasi-periodic signals, but also aperiodic and noise-like components may appear. Such characteristics can be mapped to signal generation models, and also imply that analysis of longer-term behavior will improve the performance of compression algorithms. The importance of periodic components is related to the harmonic properties of many audio sources; harmonic spectra appear in cases of voiced speech and many types of musical instruments, which can directly be related to the physical process of sound generation. Most speech coding schemes are based on linear prediction methods. Synthesis models for speech generation have also reached an advanced status and can be used for parametric coding. In the sector of speech, understandability of the reconstructed signal is an important design criterion. In general-purpose waveform coding of audio signals, utilization of psycho-acoustic properties of hearing is a key factor to achieve what transparent quality, compression without noticeable difference as compared to a digital (PCM) original. The typical core of these methods is transform coding based on block-overlapping or subband transforms. Sound synthesis, enabling synthetic generation of natural or newly-created sounds, is a very mature technology as well. It is however mostly based on nonlinear models, universal usage in analysis/synthesis audio coding is still an area requiring further development.