In the last couple of years many vendors offer noise cancellation software that is aimed to attenuate the ambient noise in phone calls. The first question that comes into mind when testing such a software is, does my peer hear the ambient noise or not? Can he/she hear me slamming the door? What about my dog barking or the car horn from a nearby street? At a first glance you would say that if the answer to any of these questions is positive then the noise cancellation is not working as expected and you should continue looking for a better solution. But is this really the case? In this post we will highlight some additional factors that should be taken into account when comparing the quality of different noise cancellation software. Having a sparking whiskey sniffer is a good goal but sometimes over brushing might be a bad idea.

Voice quality and MOS

The main purpose of the noise cancelling software is to improve the intelligibility of the voice and make the call more pleasant. As we know, noise is a nuisance and reduces the intelligibility of the voice therefore removing it should help to increase the quality of the call. But, what if while aggressively removing the noise also the voice quality is degraded, what if during noise removal process the voice becomes robotic, muffled or maybe even small portions of it are cut? Surely there is a balance between the amount of noise removed and the allowed degradation to the voice quality. The goal is to maximize the quality of the call (a.k.a. MOS score) and sometimes keeping some of the noise and not completely eliminating it will sound much better than removing all the noise while degrading the voice. In one of our previous posts we discussed an opposite technique in which noise is added in order to improve the quality of the call.

Performance

In addition to overall call quality, there is another technical factor that should be addressed. How heavy is the noise cancelling software? Does it require significant amount of CPU/GPU power and will run down your battery and will cause other applications including the communication software to lag? You are invited to read the following post regarding CPU considerations.

Distortion

As part of noise removal, especially aggressive noise removal, the original voice might get distorted. In this section we discuss distortion that is not noticeable to the human ear but can influence the accuracy of automatic voice analytic like speech to text, and speaker identification. It is important to verify that the accuracy of automatic voice analytic engines improves, or at least not reduced, when filtering audio after it was processed by the noise removal algorithm.

Conclusion

The binary answer to the question: do you hear the noise or not is important but when comparing noise cancelling software you should also measure other parameters that are not less important.