SINGLE-ENDED PACKET LOSS RATE ESTIMATION OF TRANSMITTED SPEECH SIGNALS

In this paper, we present a packet loss rate estimation model that only requires the degraded speech signal, without the original clean reference. This allows for an easy way to measure the packet loss rate of "black box" speech communication systems, for which we cannot obtain information about internal network parameters. It makes setting up complex measurement scenarios unnecessary, since any sentences of any speaker can be used as a test signal. In modern voice networks, packet loss is one of the main quality impairments, as the transmission is conducted digitally. An estimation of the packet loss rate therefore gives a helpful indication on whether an unsatisfying speech quality is caused by the network or by other factors, such as the terminal device. To detect lost packets in the speech signal we calculated MFFCs and trained a random forest. The classification results were then used to estimate the packet loss rate. The model was trained and tested on three large databases with overall 246 different speakers and 984 semi-spontaneous sentences.