Applying Hilbert-Huang Transform To Speaker Recognition

With the rapid development of the information technology and communication technology, individual identification becomes more and more important. As a biological authentication technique, speaker recognition has many applications for its advantages such as convenient, safer, and unforgotten. Speaker recognition is to analyze the feature parameters which are extracted from series of speech signals, and to identify or verify who spoke them.Hilbert-Huang transform (HHT), its key part is empirical mode decomposition, is a newly signal processing method proposed by Huang in 1998. In this thesis, we apply HHT to speaker recognition and obtain some initial achievements. In the first part, we introduce the sketch of speaker recognition and the HHT, in the other parts, the main works are discussed.We propose using linear prediction to resolve the endpoint effect evoked in EMD. Applying linear prediction to extend the data sequence by adding some predicted points, and it is stopped when a new local maximum and minimum points are emerged. Thus the error is reduced occurring in the original end when splining envelop of the data. By comparing this method with the others, we find it is effective when dealing with the endpoint effect of the EMD.On the basis of EMD, six types of parameters are presented to describe the individual features of the speaker. These parameters are EIF, ED, IMF-MFCC, IMFMFCCW1, IMFMFCCW2 and CEI, respectively determined through different methods. Then, valid test is performed by calculating the D ratio to judge their effectivity, from the results we know CEI among the six features is the most effective.At last, we perform some speaker identification experiments based on vector quantization of the above mentioned feature parameters respectively. As is shown in the results, of the six parameters, CEI has the best effectiveness as a individual feature of the speaker, and then IMF-MFCC, IMFMFCCW2, IMFMFCCW1, while EIF and ED are not suitable to be parameters for speaker recognition alone. According to a...