Chung Lee

Impact of MP3-compression on timbre space of sustained musical instrument tones

MP3 compression is widely used in music sharing and storage. A number of studies have investigated the discrimination of instrument tones after MP3 compression. Additionally, a number of previous studies have evaluated other data reduction methods including frequency modulation (FM) synthesis, wavetable synthesis, and principal component analysis (PCA). However, these studies have not considered the impact on timbre space after data reduction. In this study, listening test subjects were asked to rate the dissimilarity of all pairs of original instrument tones. The same process was done on MP3-compressed tones with various bit-rates. Correlation analysis was done on the dissimilarity data of the original and compressed tones to see if MP3 compression caused a significant impact on the perceptual distance between instrument pairs. The multidimensional scaling (MDS) solutions of the original and compressed tones were also compared to see if the timbre space was significantly altered after MP3 compression (e....

Discrimination of MP3‐compressed musical instrument tones.

Instrument tones were compressed using an MP3 codec to determine how the detection of compressed sounds varies with bit rate and instrument. Listeners were asked to discriminate the compressed sounds from the original sounds. Discrimination was very good for low bit rates, moderate for intermediate bit rates, and poor for higher bit rates. Discrimination varied significantly from instrument to instrument. Discrimination scores were strongly correlated with the signal-to-mask ratio (SMR) and spectral irregularity of the original tone. Relative spectral error accounted for more than 80% of the variance in the discrimination scores.

Discrimination of instrument tones resynthesized with piecewise‐linear approximated harmonic amplitude envelopes.

Harmonic amplitude envelopes of instrument tones were approximated using piecewise‐linear segments to investigate how the discrimination of approximated sounds varies with the number of breakpoints and instruments. Listening tests were done to determine the discrimination of the approximated sounds from the originals. Apart from basic statistical analysis, various error measurements were evaluated based on the results of the listening tests. Relative‐amplitude critical band error was the best error metric, accounting for about 82% of the discrimination variance. Strong correlations were found between the discrimination scores and spectral incoherence, with breath noise in the flute and bow noise in the violin particularly discriminable in piecewise‐linear approximation.