Voice over Internet Protocol (VoIP) networks are an increasingly important field in the world of telecommunication due to many involved advantages and potential revenue. Measuring speech quality in VoIP networks is an important aspect of such networks for legal, commercial and technical reasons. The E-model is a widely used objective approach for measuring the quality as it is applicable to monitoring live-traffic, automatically and non-intrusively. The E-model suffers from several drawbacks. Firstly, it considers the effect of packet loss on the speech quality collectively without looking at the content of the speech signal to check whether the loss occurred in voiced or unvoiced parts of the signal. Secondly, it depends on subjective tests to calibrate its parameters, which makes it applicable to limited conditions corresponding to specific subjective experiments. In this paper, a solution is proposed to overcome these two problems. The proposed solution improves the accuracy of the E-model by differentiating between packet loss during speech and silence periods. It also avoids the need for subjective tests, which makes it extendable to new network conditions. The proposed solution is based on an Artificial Neural Networks (ANN) approach and is compared with the accurate Perceptual Evaluation of Speech Quality (PESQ) model and the original E-model to confirm its accuracy. Several experiments are conducted to test the effectiveness of the proposed solution on two well-known ITU-T speech codecs; namely, G.723.1 and G.729.