As I know the MOS scale is subjective and is prepared by ITU based on thousands of people listenning to samples. These vendors develop their own algorithms which map these samples to the MOS scale. They are quite aquarate. These speech samples are specially designed for different languages and contain male and female voices and etc. Then the reference sample is send from the one end (ex. MS) to the other (ex. ISDN) and then again compared with the reference. According to the algorithm they are maped from 1 (bad) to 5 (excellent). For example if I remember the highest possible value with ISDN to ISDN is is 4.2 and GSM EFR to ISDN around 3.9.
What do you mean with golden ear?