Audio quality assessment in packet networks: an "inter-subjective" neural network model

Transmitting digital audio signals in real time over packet switched networks (e.g. the Internet) has set forth the need for developing signal processing algorithms that objectively evaluate audio quality. So far, the best way to assess audio quality are subjective listening tests, the most commonly used being the mean opinion score (MOS) recommended by the International Telecommunication Union (ITU). The goal of this paper is to show how artificial neural networks (ANNs) can be used to mimic the way human subjects estimate the quality of audio signals when distorted by changes in several parameters that affect the transmitted audio quality. To validate the approach, we carried out an MOS experiment for speech signals distorted by different values of IP-network parameters (e.g. loss rate, loss distribution, packetization interval, etc.), and changes in the encoding algorithm used to compress the original signal. Our results allow us to show that ANNs can capture the nonlinear mapping, between certain characteristics of audio signals and a subjective five points quality scale, "built" by a group of human subjects when participating in an MOS experiment, creating, in this way, an "inter-subjective" neural network (INN) model that might effectively "evaluate", in real time, the audio quality in packet switched networks.