INVESTIGADORES
DI PERSIA Leandro Ezequiel
congresos y reuniones científicas
Título:
Validity Studies of Blind Source Separation Using Independent Component Analysis in Real Environment
Autor/es:
KENSUKE TSUTSUMI; AKIHIRO HANDA; LEANDRO DI PERSIA; MASUZO YANAGIDA
Lugar:
Kyoto
Reunión:
Congreso; Meeting of the Acoustic Engineering Society; 2004
Institución organizadora:
Institute of Electronics, Information and Communication Engineering
Resumen:
Investigated in this report is validity of frequency domain Blind Source Separation using Independent Component Analysis in real environment. Test sounds were recorded in a sound-proof chamber and in a living room simulator. Then frequency domain Blind Source Separation was applied for the sounds. FastICA is used as a learning algorithm with theinitial value produced by JADE. The effect of separation is evaluated by improvements in S/N and speech recognition rate. Also investigated is how we can maintain the separation performance even in bad situations. The result shows that speech recognition rate is improved about 10% in environment having reverberation of about 150ms.