Subjective evaluation of quality and intelligibility of speech distorted by synthesized noise

Abstract

The distortion of the speech signal by noise interferences negative impacts on the perception of speech information by listeners, and a noise disturbance in the form of people conversations has the best masking ability. This phenomenon is usually used when the intelligibility of speech should be minimal. Therefore, there are nowadays many different acoustical systems generating acoustic noise in the form of stationary or non-stationary noise for active masking of speech information. Assessment of acoustic masking quality for systems generating stationary noise can be made by means of formant technique and speech intelligibility can be used as a measure of masking quality for such systems. Previously, it was theoretically shown that masking property of white noise is worst at low signal-to-noise ratio. However, this result was not tested by subjective testing. Moreover, masking ability of nonstationary noise was not tested too. In this paper, this gap has been eliminated and the results of subjective estimation of the effect of stationary and nonstationary synthesized noise on the quality and intelligibility of speech signals are presented. Degradation Mean Opinion Score (DMOS) measure of speech quality was used for the estimation. It was used the fact of high correlation (about 0.9) between speech quality and intelligibility upon results interpretation. For the case of stationary noise, it was shown that for signal-to-noise ratios below minus 5 dB white noise is inferior to pink and brown noise by masking ability. This result is in a good agreement with previously theoretically predicted one. Two simple, from the computational point of view, non-stationary noise generation algorithms were studied also. The first algorithm uses both spectrum inversion and reverberation simulation. Second algorithm is based on formation of nonstationary process as result of noise carrier amplitude modulation by envelope of speech signal. It was found that these nonstationary processes provide a better, in comparison with white noise, masking of speech signals. These nonstationary processes have the significant advantage compare to stationary ones because they provide less environmental pollution during speech pauses.Ref. 12, fig. 8.