Research On MNB2 Algorithm And Objective Evaluation Platform Of Speech Codecs

With the rapid development of information technology in communication field, many speech codecs have been emerging. In order to measure the quality of these speech codecs, researchers have built many subjective listening test plans to examine all aspects of a speech codec. The prime advantage of subjective evaluation is that its result represents the real perception of human. However, subjective listening test is time consuming, labor-intensive, expensive, inflexible, nonrecurring and unstable. So we need the objective evaluation to avoid the disadvantages of the subjective listening test. The study of objective evaluation can greatly improve other relative research field, such as speech recognition, speech enhancement, and so on.Firstly, this dissertation fully summarized subjective evaluation methods and objective evaluation methods, and demonstrated the features of each objective methods. In this dissertation, I studied the world's most advanced I/O-based objective evaluation algorithm~MNB2 Algorithm, which combines the delectability model and the judgment model and totally improves the model built by early researchers, In my research, I studied the psychoacoustics principle of this algorithm and a concrete structure for telecommunication band speech codec. And then, I simulated the algorithm in MALAB. During the simulation study, over 1,500 speech segments, about 11s each, were processed more than 5,000 times and got about 4,000 data. The accuracy of the algorithm, the difference between the evaluation results of different languages and the background noise's effects to the evaluation, were all included. The results demonstrated that the algorithm is very meaningful and can be applied to the evaluation of new speech codec. And a reduced MNB2 model that got from early research was also examined. The results showed reduced model and complete model could get almost the same evaluation data. At the same time, the reduced model can save many memories.Finally, A C++ implementation of this algorithm was integrated into a objective evaluation platform of speech codec. OOP(Object Oriented Programming), UML(Unified Modeling Language), DirectX and multithread techniques were used in the development of this platform.