Subjective evaluation of four low-complexity audio coding schemes

Abstract

In this study the subjective performance of four low-complexity audio data compression methods are compared, operating at nominal bit rates of 2, 3, 4, and 5 bits per sample, applied to four 20-kHz bandwidth, 16-bits per sample digitized musical signals. The simple compression schemes compared were elementary differential pulse-code modulation (DPCM), noise feedback coding DPCM (NFC-DPCM), adaptive quantizer DPCM (DPCM-AQB), and a recently proposed method known as recursively indexed quantizer DPCM (RIQ-DPCM). Pairs consisting of a reconstructed signal and a reference signal were presented in a two-interval preference experiment. The reference signals were processed for specified levels of modulated noise reference unit (MNRU) in order to estimate the equality threshold rating (ETR) of the reconstructed audio stimuli. The subjective MNRU values were found to increase by 2-5 dB for each increment in bits per sample. The DPCM-AQB scores were found to be 8-10 dB higher than for DPCM and NFC-DPCM. RIQ-DPCM was rated highest, exceeding the DPCM-AQB results by 2-5 dB in all tests. Objective measurements of segmental signal-to-noise ratio (SNRSEG) for the reconstructed signals predicted a performance level 2-5 dB lower than was actually found in the subjective results, particularly for SNRSEG values below 25 dB.