| Summary: | 碩士 === 國立中正大學 === 電機工程所 === 97 === Independent component analysis is widely adopted, especially in blind source separation, data clustering, speaker recognition and parameter extraction. This dissertation mainly explores to separate 2 and 4 sound sources from the recorded signals. The mixing signals from 2 and 4 microphones are analyzed by using blind source separation. The Probability Density Function（PDF）associated with the Minimization of the Mutual Information (MMI) is adapted according to the correlation and characteristics of the mixing signals. The correlation of the separated sources is investigated to determine the adequate PDF from the super-Gaussian function and kernel density function. Additionally, separated sound signals are classified into speech, audio, noise and so on in which one of the super-Gaussian function and kernel density function is adequately selected to conduct MMI rather than the Gaussian function. Such an approach allows our prediction model closer to the PDF of the original signals. In our experiments, different kinds of sound signals with different correlations are employed to verify our adaptive model. The proposed adaptive model selecting the adequate PDF can effectively improve the correctness of sound source separation. The SIR values are improved around 2.5 and 1.0dB in average for the situations of 2 sources to 2 microphones and 4 sources to 4 microphones, respectively. Therefore, the proposed adaptive model used in MMI of blind source separation can be widely applied to various independent component analyses.
|
|---|
