Adaption of a Prediction Model for Noisy Speech Quality Assessment

Abstract:
The objective model according to ETSI EG 202 396-3 [1, 2] was developed for predicting speech, noise and global quality of noisy speech signals for wide- and narrowband terminals according to ITU-T recommendation P.835 [3]. Today’s signal processing capabilities of modern terminals (mobiles, smartphones) have been rapidly developed in the last years, even 2-channel-microphone noise reduction solutions are currently state of the art. In result, a much higher speech and noise quality can be achieved with these devices than in the past. This work describes a way to adapt the narrowband mode of the model to this new high quality range which could not be covered within the development in 2007 (narrowband extension of [1]) also taking into account a higher SNR ranges. The work conducted shows that only minor adaptations of the method described in [1] are required to significantly improve the prediction performance. It can be shown that the adapted model shows an excellent backward compatibility to the old databases used for training and validation. Index Terms — objective measurement of speech quality, 2 channel noise reduction, mobile devices