Serija predavanja prof. Kleijna

Lecture 1.

Title: MODELING AND CODING OF SPEECH AND AUDIO SIGNALS

 

The use of coding for transmission is clear, but it has been argued coding can also be used as a generic measure of goodness for the model family (e.g, the autoregressive models of order 10) used in the coder. Speech and audio coding are, therefore, relevant in the selection of signal model families for subjects such as speech recognition and signal enhancement. To evaluate model families for a particular application, the coders must be optimized with respect to a suitable distortion measure. The usage of complex distortion measures is not trivial and we discuss the standard approaches and open problems in this area. To find the optimal coding rate for a model family given a distortion efficiently, it is advantageous to exploit any knowledge about bit allocation to the model selection (e.g., the coefficients of the autoregressive model commonly computed by linear prediction) and the signal segment given the model. We show that under different commonly assumed conditions the bit rate required to select the model is independent of the overall rate (and of signal distortion) and that the allocation can be computed. We find that existing practical coders satisfy the predicted relation, even though they were developed without knowledge of it. This suggests that the models used in speech and audio coding are indeed useful for other applications and that the bit allocation step can be eliminated.

Lecture 2.

Title: SOURCE CODING FOR HETEROGENEOUS NETWORKS

 

The usage of packet networks for voice traffic has introduced challenges to the transmission of audio-visual information that differ significantly from those encountered in traditional circuit-switched networks. The lower cost of packet networks is associated with long delay, bit errors, and packet loss, each varying in severity with the networks used, over time, and with the application. We provide an overview of technologies that can be used to make the transmission of audio-visual information efficient over heterogeneous packet networks. We describe in more detail two methods that are of particular significance. First, we describe high-rate coding theory, which leads to analytic methods for coder design. As a result, source coders can be redesigned in real-time and can adapt to the network conditions encountered by a particular communication service at a particular time. Second, we discuss multiple description coding (MDC), which distributes signal descriptions over different packets. The signal can be reconstructed from each individual encoding. What makes MDC special is that the quality of the reconstructed signal increases with the number of descriptions received.

 

Lecture 3.

Title: SIGNAL ENHANCEMENT FOR SPEECH AND AUDIO

 

This talk provides an overview of methods commonly used for the enhancement of single-channel signals contaminated by additive noise. We use a systematic statistical modeling viewpoint. In most existing paradigms, the clean signal is estimated under the (somewhat unreasonable) assumption that the statistics of the clean signal and of the noise are known. We first discuss clean-signal estimation methods under this assumption. We include historic methods and the more formal maximum likelihood (ML) and minimum mean-square error (MMSE) estimation methods. Usually the clean-signal and noise statistics are described in the form of simple power spectra or covariance matrices, but Gaussian mixture and Markov models are increasingly common. Next, we discuss methods to estimate the noise and clean-signal statistics. We include the quantile, minimum statistics, and ML and MMSE estimates. Finally, we show that it is possible to formulate the problem of estimating the clean signal without making a specific selection for the noise and speech models. For certain cases, this general formulation leads to existing methods. We note that the general approach does not guarantee results that are physically reasonable and conclude that enhancement remains an art rather than a science.