Extraction of acoustic signals using blind source separation method.

Abstract

This paper presents the results of acoustic feature extraction using blind source separation. The objective of this study is to separate individual acoustic signals based on measurements of overall signals without knowing the number of sources, their locations, and how signals are mixed. The only requirements are (1) the acoustic sources are statistically independent, namely, knowledge of one source gives no information on that of the other, (2) signals are non‐Gaussian, and (3) the number of sources is no more than that of microphones. Numerical simulations are conducted based on input data collected by four microphones in a free field, and results are obtained using fast independent component analysis (ICA) in time domain. Different types of acoustic sources, including human speech,music, impulsive sounds, machine noise, helicopter sounds, etc., are used. The impacts of microphone spacing and locations, source locations, signal to noise ratio, sampling rate, and various methodologies such as maximization of non‐Gaussianity, maximal likelihood estimation, minimization of mutual information, etc., on separations of individual sources are examined. Results show that different approaches lead to different levels of success in source separation. In general, satisfactory results can be obtained when sources are spatially isolated and reverberation effects are negligible.