Feature Extraction Method Of Noisy Speech Based On EMD And Feature Normalization

In recent years, with the rapid development of the mobile internet, speech recognitiontechnology which tried to replace physical keyboard computer now has been successfullyapplied to all kinds of mobile terminal, and indeed changed people’s life. However, since thespeech recognition technology was proposed, the robustness identification of the system withnoise has always been the hot and difficulty issues in the study, and it is also a great obstaclewhich limits the speech recognition technology’s development.As everyone known, the speech recognition system can be a fairly ideal identificationaccuracy in an ideal environment. However, in our life there is environmental noise andchannel convolution noise everywhere, it would cause a sharp decline of the accuracy ofspeech recognition system. This is mainly because the characteristic parameters of the inputsignal don’t match with that of the trained data in the noise environment identification system.According to this, the paper mainly studied the feature extraction method of speechrecognition in the noise environment. Upon the previous research, we did effort in the signalspace robust method and in the feature space robust method. The specific work and researchachievement are as follows:Firstly, summarized the development history and present situations of speech recognitiontechnology in the first part of this paper; compared with the current noisy speech recognitiontechnology, stated the topic’s background and meaning.Secondly, because of the diversity of the actual noise, many methods were produced inrecognizing the robustness of speech recognition system. at the same time, the realization ofautomatic identification system is based on the statistical modeling theory, therefore, it firstintroduced the automatic speech recognition based on statistical modeling, then theapplication of signal space robust method, feature space robust method and model spacemethod in speech recognition are introduced as the theoretical basis of the subsequentchapters.Furthermore, it mainly introduced empirical mode decomposition (EMD) algorithm insignal space method. First is the application in the field of speech enhancement, and from theexperimental we could know the method is good at the noise suppression. Then proposed that apply it in the pre-treatment of speech recognition and put forward the noise environmentbased on EMD robustness speech feature extraction algorithm. Finally, verified theeffectiveness through the theoretical analysis and the experiments.In the fourth part, inspired by features transformation algorithm such as cepstrum meannormalization (CMN) algorithm and cepstrum variance normalization (CVN) algorithm, triedto propose speech feature extraction method based on the power spectral density (PSD)normalization, which is based on the cepstrum mean variance normalization (MVN)algorithm from the statistical distribution of the speech feature. By the experiment we couldmake the conclusion that the proposed feature extraction method based on PSD algorithmcould further suppress the environmental noise in feature space, reduce the influence of thedifferences of the characteristics parameters between test data and trained data, and improvethe speech recognition system identification accuracy in some degrees. At the same time, thismethod is simple, small amount of calculation, so it is easy to implant in the applicationsystem.Finally, summarized all the above, and the existing problems of this method were given,and these problems would be studied as the next research emphases.