Model-Based Dereverberation for Speech Recognition

Inhalt:
This paper presents an effective method for suppressing reverberation and noise for speech signals. Based on a simple model which considers basic principles of room acoustics the power spectral density of the reverberation is estimated. Spectral subtraction is then used to weight the spectrum of the input signal dynamically. The method is characterized by its robustness and efficiency. It enhances speech recognition performance considerably for singlechannel as well as for multi-channel systems.