Error Concealment and Error Protection for Audio over IP Networks
MetadataVis full innførsel
Audio streamed over current packet switched Internet Protocol (IP) networks may be distorted due to packets arriving after their scheduled playback time and even packets lost in the network. To keep time synchronization and minimize the resulting distortions, the receiver has to perform some form of Error Concealment (EC). The sender can in addition add Error Protection (EP) to reduce the effect of missing packets at the cost of a higher bit rate. This thesis consists of three papers. In paper A an EP scheme is presented. The sender simulates a packet loss for each audio frame, and evaluates by using a perceptual model how well the EC algorithm performs compared to a low bit rate encoded version of the same frame. If the low bit rate frame performs a given amount better than the frame estimated by the EC algorithm, the low bit rate frame is sent as EP. Using this scheme, one can adjust the threshold for when EP is used, and the quality improvement vs. redundancy can be quantified. Two listening tests have also been conducted. In paper B, distortions in audio signals due to packet loss are evaluated. Three different EC algorithms, four packet loss rates and five clips were evaluated by a total of 21 non-expert listeners. The test results show significant differences between packet loss rates, EC algorithms, and also audio clips. A two-parameter model for the perceived quality as a function of packet loss rate is also presented. Effects of network architectures are investigated in the third paper. Models for packet loss processes for both a Best Effort (BE) router and a Differentiated Services (DiffServ) enabled router were found using network simulations. Both models were combined with three packet loss rates, three audio clips, and two EC algorithms, and the perceived audio quality was evaluated. 24 non-expert listeners participated in the subjective test. The results show no significant differences between the two loss models, but interaction effects between packet loss model and both audio clip and packet loss rate can be seen.