Discrimination of Data-Reduced Sustained Musical Instrument Tones

We know that musical instrument tones are recognizable even if they are altered. The current study investigates the perception of musical instrument tones altered by two data reduction methods: MP3 compression and piecewise linear approximation (PLA) of additive synthesis amplitude envelopes. Sustained musical instrument tones were data-reduced using the above methods to determine how the detection of data-reduced tones varies with instrument and the degree of data reduction. Sounds with harmonically-flattened frequencies were compressed by MP3 compression and PLA of amplitude envelopes. Listeners were asked to discriminate the altered sounds from reference sounds resynthesized from the original data. This allowed us to determine which degree of data reduction produces near-perfect, (above 90%), moderate (around 75%), and poor discrimination (around 50-60%). Statistical analysis showed that discrimination was different from instrument to instrument. Discrimination scores were strongly correlated with a number of spectral measurements of the original tone. Objective error metrics including relative spectral error were compared for their correspondence with the discrimination scores. Other than discrimination, multidimensional scaling (MDS) was done to search for salient timbral attributes other than spectral centroid and attack time. In addition to that, in follow-up work we plan to do a study where listeners will be asked to rate the dissimilarity between the MP3-compressed instrument tones. MDS solutions of the compressed instrument sounds and the originals will be compared to determine if MP3 compression causes significant impact on instruments’ relative positions within the timbre space.