Towards a Better Understanding of the Effect of Reverberation on Speech Recognition Performance

In order to tailor dereverberation approaches to automatic speech recognition (ASR) systems, it is important to thoroughly understand the effect of reverberation on ASR performance. In this work, the effect is analyzed by varying the shape of the room impulse response (RIR) using two design parameters that are useful for defining the target response of typical dereverberation algorithms. The parameters determine the amount of attenuation of the coefficients that correspond to reflections arriving with at least a delay of T after the direct-path component. By convolving clean speech signals with the modified RIRs, ideal late-reverberation suppression is simulated. By varying the level of attenuation A and the delay T , the influence of these design parameters on the recognition rates is investigated. Thus, guidelines for adjusting dereverberation algorithms to ASR systems are deduced.