Separation of speech signal - To realize multiple talker speech recognition

The rapid advances in automated speech recognition technology has enabled computers to recognize human speech with a high accuracy, if the speaker speaks politely into a microphone close to the mouth. However, the recognition rate decreases considerably when there are obstructive sounds such as another person's voice, background music, ambient noise, or reverberation. In such cases, computers are unable to recognize what was said. Recently, a statistical method called Independent Component Analysis (ICA) has attracted the attention of researchers as a technique for sound source separation. On the assumption that the sound sources, that is, one's voice, another person's voice, background music, and so on, are mutually independent, this method can restore the original signals if the observed signals are separated to statistically independent signals. This is the principle of source separation using ICA. Our approach and current results are shown in this paper.