Colloquium - Details

Master-Vortrag: Self-organizing Prediction for Multi-Channel Audio Coding

Niklas Koep
27. Januar 2014
12:00 Uhr
Hörsaal 4G IKS

In this thesis, the concept of linear predictive coding for multi-channel audio signals is examined. While predictive audio codecs mainly focus on coding of individual channels, this work focuses on methods geared towards the exploitation of preexisting correlations across different channels.

A generalization from linear predictive single-channel audio coding to multi-channel signals is presented in which the prediction signal of each channel is calculated from all available channels of a multi-channel signal. It is shown that systems with different filter orders for each prediction filter generally perform similar to prediction setups that use a fixed order for every filter in the system. The latter constitutes a particularly important configuration in practice due to its close relation to multivariate autoregressive models and because it is the only case for which stability of the corresponding synthesis filter is guaranteed. A generalization of the well-known Levinson-Durbin algorithm for efficiently solving the multi-channel normal equations is derived for this setup.

The influence of quantization of the prediction error on the reconstructed signal is examined. A novel approach for spectral shaping of the multivariate quantization error is proposed which restrains the influence of quantization noise to the corresponding reconstructed channel.

A novel parametrization of the analysis filter based on the polynomial eigenvalue problem is proposed for the quantization of the filter coefficients which exhibits similar properties to the line spectral frequencies. The method encodes the matrix polynomial in terms of its latent eigenvalues and what is termed the pseudo-eigenvectors as obtained from the solution of the associated generalized eigenvalue problem.

back