Speech-model processes controlled by discrete Markov-chains

Up to now, the general problem of speech quality assessment, especially of coded speech signals has been accomplished by using highly standardized human listening tests. Because of several shortcomings of these subjective procedures many objective approaches have been proposed in the past. Each of the objective measures tries to avoid most of these drawbacks. However, one of the major problems is still left out of account by all of them. The test results depend on the underlying set of speech-data (e.g. speaker dependence!), since they use natural speech as the test signal. So, in this paper we propose a test signal, which is a special random process. It is an extension of a known model in terms of slowly time-varying features. It is shown that the important characteristics of this process agree with those of natural speech. The time-variant properties are controlled by trained Markov-chains to achieve a speech-model process with natural formant- and pitch-structures. This process is described and some measurements results are given