Subjective evaluation of audio compression algorithms: Do conservatory students have a ‘‘golden ear?’’

Abstract

Digital audio compression algorithms are typically based upon psychoacoustic models of hearing; thus they are known as perceptual coders. For commercial applications such as providing music over the Internet, where bit-rate reduction is essential, these lossey codecs have customarily been evaluated through subjective testing using professional audio engineers as listeners. This study replicated the MPEG-2 AAC Stereo Verification Tests (ISO/IEC JTC1/SC29/WG11 N2006) with the 20 listeners drawn from undergraduates enrolled in the Oberlin Conservatory of Music. These tests used the triple-stimulus hidden-reference double-blind methodology with eight codecs and 10 test stimuli. In this procedure, listeners compare two test signals (one coded, the other not) to an uncoded reference, then rate the ‘‘audio quality’’ of both test stimuli on a 1–5 scale in 0.1-step increments. The test stimulus selected as the hidden reference is graded 5.0, the coded signal graded on the degree of impairment. Difference scores thus reflect the discriminability of the stimuli. Taking the magnitude of Studentized t-statistic values as a measure of acuity, MPEG listeners averaged 4.45, while Oberlin listeners averaged 9.86. While there were some methodological differences between the two experiments, subjects differed several ways Oberlin listeners were younger and were engaged in formal musical training.