Typically, Noise Suppression Algorithms (NSA) operates on a noisy speech waveform and attempt to reduce the noise or background component without adversely affecting the speech or signal component of the waveform. This goal can often be realized for relatively low levels of noise suppression. For higher levels of noise suppression, however, NSA's often adversely affect the speech component as more noise is suppressed –there tends to be increasing degradation of the speech or signal component as more of the noise or background component is removed. In this situation subjects can often become confused as to what they should be responding to in their ratings of the overall “quality” of the waveforms – while the background may have been improved because there is less noise present in the waveform, the speech signal may have been degraded in the process. In a single-scale rating method, the ACR for example, each individual subject weights the signal and the background components in determining his ratings of overall speech quality. This weighting process introduces additional error variance in the subjects ratings of overall quality resulting in decreased reliability in those ratings. The methodology described in this Recommendation reduces the listener’s uncertainty by requiring him to successively attend to and rate the waveform on: the speech signal, the background noise, and the overall effect-speech + background.
ITU-T P.835 FRENCH-2003 history
2003ITU-T P.835 FRENCH-2003 Subjective test methodology for evaluating speech communication systems that include noise suppression algorithm