Improvement of speech signal extraction method using detection filter of energy spectrum entropy

In speech recognition system research, recognition system performance has been significantly improved through research and development in the speech recognition area, but environmental noise is still a favorite subject for research due to its numerous environmental changes. And speech extraction techniques, which are widely applied, improve speech signals that are mixed with noise. A least mean square (LMS) adaptation filter is commonly used to help noise estimation and detection algorithms adapt to changing environments. But an LMS filter needs some time to adapt and estimate signals. That weakness can be overcome by using energy spectrum entropy and an average estimate LMS (AELMS) filter to detect robust voice activity in a noisy environment. In this paper, we propose a speech signal extraction method using a detection filter of energy spectrum entropy. The proposed method is polluted speech–signal noise extraction to reduce noise with an AELMS filter to detect robust voice activity. An AELMS filter maintains source features of speech, decreases speech information degradation, and reduces noise in a polluted speech signal. To improve adaptation speed, we calculated an average estimator, and controlled the LMS filter step size with a frame measure. For speech detection of signals synthesized with low-speed and high-speed driving noise, an energy spectrum entropy method was used. Compared to an existing method of using frame energy, the proposed method improved the starting point of the resulting speech by 1.7 % of an error rate and by 3.7 % of an end point error rate.