Audio Signal Processing and Coding

12.7 ALGORITHMS FOR PERCEPTUAL MEASUREMENT

Numerous examples of both CIR- and NSE-based perceptual measurement schemes for speech and audio codec output quality evaluation have appeared in the literature since the late 1970s, with particular emphasis placed on those systems that can predict accurately subjective listening test results. In 1979, Schroeder, Atal, and Hall at Bell Labs proposed an NSE technique called the “noise loudness” (NL) metric for assessment of vocoder output quality. The NL system [Schr79] estimates output quality by computing a ratio of the coding distortion loudness to the loudness of the original signal. For both signal and noise, loudness patterns are derived every 20 ms from frequency-domain excitation patterns that are in turn computed from short-time critical band densities. The critical band densities are obtained from FFT-based spectral estimates. The approach is hampered by insufficient analysis resolution at low frequencies. Karjaleinen in 1985 proposed a CIR technique [Karj85] known as the “auditory spectrum distance” (ASD). Instead of using an FFT for spectral analysis, the ASD front-end models the cochlear filter-bank using a 48-channel bank of overlapping FIR filters with roughly 0.5 Bark spacing. The filter bank channel magnitude responses are designed to model the smearing of spectral energy that produces the usual psychophysical excitation patterns. Then, a square-law post-processor and two low-pass filters are used to model hair cell rectification ...