Blind Separation for Real-World Speech Signals

Among potential applications of blind source separation (BSS) a most promising one might be separation of speech signals. In real-world situations, however, any BSS algorithm for sound signals suffers from a difficulty. From a practical point of view the microphone array should be made compact, but then the mixing matrix becomes almost singular, inducing certain instability in the algorithm execution. This paper describes some experiments of BSS, which were made in a soundproof room, an office room, and a car. The results show that an appropriate configuration between the microphone set and the sound sources is very important to achieve satisfactory separation. Moreover, astonishingly, if the microphones are located appropriately, even using only two microphones considerably enhances a target sound from mixtures of more than two sounds. Key-Words: blind separation, independent component analysis, speech, voice, noise cancelling, minimal distortion