DFKI-LT - Detecting Packet-Loss Concealment Using Formant Features and Decision Tree Learning
Detecting Packet-Loss Concealment Using Formant Features and Decision Tree Learning
1 Proceedings of Interspeech 2018,
One of the main quality impairments in today's packet-based voice services are interruptions caused by transmission errors. Therefore, most codecs comprise concealment algorithms that attempt to reduce the perceived quality degradation of missing speech packets. In case the algorithm fails to properly synthesize the lost speech, interruptions or unnatural sounds are usually perceivable by the user. When measuring the quality of a voice network, there are excellent tools available, which can predict the perceived speech quality. However, they offer only little insight into the technical cause of a quality degradation. A packet-loss detection model could explain the influence of transmission errors on the speech quality and state a packet-loss rate. Thus, making it easier to identify technical problems in the network. In this paper, we examine a new approach for detecting (perceived) packet-loss of transmitted speech by audio analysis. After finding a lost packet, the model classifies in a second stage if the loss was perceivable as a quality degradation. In the model, we use meaningful features that are easy to interpret, and obtained promising results in a simulated environment. Therefore, this detector could also be used to evaluate new packet-loss concealment algorithms and help in optimizing the same.
Files: BibTeX, Interspeech.2018-1098, external_mittag_2018_detecting-packet.loss-concealment-using-formant-features-and-decision-tree-learning.pdf