Sound Category Detection Including the Human Speech like Noise after Independent Component Analysis and the Sound Trace
Kazuhide Okada · Procedia Engineering · 2012
For separating the mixed sound into the each separated sounds including the noise removing, ICA (Independent Component Analysis) is frequently used, and the Kalman Filter is used for the sound trace and then elimination of the sound itself is done. But even after the separation of the sound by ICA we cannot yet say which category the said sound belong to or even if the specified sound is traced by Kalman Filter, whether its sound is the target which should be taken out as the pure one or ot cannot be known. So the classification of the said sound is required as the post-processing, especially judging for HSLN (Human Speech Like Noise) which means the actual voice and the others. Limiting its sound pressure level of HSLN into 50 dB and that of the other sound source into 30 dB as the actual noise circumstances,the specification of the separation by ICA could be done successfully by 78.5%. By this trial Speech recognition ratio was improved from 55% to 63% averagely.