SUMMARYIn this paper, we propose frequency assignment as a method for multichannel speech separation. In this method, the magnitudes of the absolute values of each component in the frequency domain are compared, then based on this comparison result determines to which channel a frequency component originally belongs. The speech is separated by assigning the frequency to the determined channel. Therefore, it becomes possible to separate multichannel speech with a low computational load. However, it has the restriction that each speaker must speak close to his microphone. For situations in which this premise does not hold, we present a solution that uses beam-forming toward the sound sources. We also show the application of the frequency assignment to sound source localization. In this paper, we demonstrate the usefulness of the proposed method through simulation tests and tests in a realistic environment.

No abstract

A solution to the problem of improving robustness to noise in automatic speech recognition is presented in the framework of multi-band, multi-SNR, and multi-path approaches. In our word recog