Using linear interpolation to improve histogram equalization for speech recognition

This paper presents a novel approach to speech data normalization by introducing interpolation for histogram equalization. We study different ways of histogram interpolation that inhence this normalization technique. We found that using a special weighting factor to combine current and past test sentence statistics improved speech recognition performance. For the testing that used weighted histogram interpolation we achieved 44.85% phone error rate against 49.42% phone error rate for the testing without normalization and 48.59% phone error rate, when only a single test sentence histogram was used for normalization. Recognition experiments were conducted on speech data recorded in a moving car and proved advantage of using interpolation for data normalization by histogram equalization.