Target speech feature extraction using non-parametric correlation coefficient

Abstract

Speech recognition systems for the automobile have a few weaknesses, including failure to recognize speech due to the mixing of environment noise from inside and outside the car and from other voices. Therefore, this paper features a technique for extracting only the selected target voice from input sound that is a mixture of voices and noises. The feature for selective speech extraction composes a correlation map of auditory elements by using similarity between channels and continuity of time, and utilizes a method of extracting speech features by using a non-parametric correlation coefficient. This proposed method was validated by showing that the average distortion of separation of the technique decreased by 0.8630 dB. It was shown that the performance of the selective feature extraction utilizing a cross correlation is good, but overall, the selective feature extraction utilizing a non-parametric correlation is better.