An Analysis of Dereveberation Algorithms on Music Signals

The acquisition of a clean audio from distant microphones is often affected by the phenomenon of reverberation. On speech signals, the reverberation affects the intelligibility. Different ways to minimize this effect on speech signals have been studied during recent years because this is important in fields such as Automatic Speech Recognition. In music, the reverberation is an important aspect to take into account considering that it adds a sense of space and therefore, the effect is not negative. However, taking control over the reverberation could be interesting in order to clean the signal or maybe to manipulate it by changing the reverberation. In this study one commercial plug-in, named as NML RevCon, has been studied and evaluated by using a set of instrumental sounds. The algorithm behind the plugin is one of the algorithms with most chance of success to achieve clean audio, and it is based on the multi step linear prediction technique. The database was built from a set of anechoic sounds, convolved with three different impulses responses and de-reverberated with the mentioned plug-in. Having the sounds in these three states was important in order to evaluate, later, the behaviour of the plug-in. Since the plugin has a set of control parameters, the de-reverberated process has been done four times for each sample with different parameter settings. The evaluation has been done by extracting the frequency-weighted segmental signal to noise ratio, from the comparison between the anechoic sample and the de-reverberated sample. The results show that the speech sound achieves better results after the plug-in process than the instrument sounds. Taking into account only the instrument samples, the best results are achieved by the viola sounds. In addition an analysis has been done of how the plugin behaves with different playing techniques. The results show that the sounds with well defined onsets, like pizzicato or staccato, have better results than the sounds like legato or detaché