Low rate speech coding and random bit errors: a subjective speech quality matching experiment

Original publisher: [Washington, D.C.] : U.S. Dept. of Commerce, National Telecommunications and Information Administration, [2009] OCLC Number: (OCoLC)708373536 Subject: Speech processing systems. Excerpt: ... Figure 1. The human interface used during the experiment. This experiment differs from many in that the speech signals to be presented on a given trial were generated on-the-fly immediately before that trial. In the worst case, the time required to do this was 750 ms. Thus, the on-the-fly generation was not perceived as an unnatural delay in the experiment by the listeners. 2.3 Listeners Sixteen listeners participated in the experiment. These listeners were a subset of a large pool of employees that were randomly selected from the U.S. Department of Commerce Boulder Labo-ratories telephone directory and invited to participate. A total of twelve males and four females participated. Three of the listeners were estimated to be in their 20 ' s, five in their 30 ' s, six in their 40 ' s, and two in their 50 ' s. Two of the females were non-native speakers of English. Listeners were not familiar with the goals of the experiment nor the speech codecs under test. Depending on the listener, the duration of the experiment ranged from 10 to 30 minutes. 2.4 Speech Quality Matching Algorithm In addition to the functions already identified, the experiment control software also implements a speech quality matching algorithm ( QMA ). This algorithm seeks to find a BER value where the CUT has the same perceived quality as the reference condition. That is, when the CUT and the reference condition are played sequentially as a pair with randomized order, the listener votes " no preference. " Thus the algorithm searches a line segment in BER space extending from 0 % to 8 % BER. This upper limit of 8 % was chosen after a preliminary listening to CUT recordings. It was chosen to be well above the range where " no preference " votes would be likely to occur. 5