Audio Signal Processing and Coding

9.7 HYBRID SINUSOIDAL CODERS

This section examines two experimental audio coders that attempt to enhance source robustness and output quality beyond that of the purely sinusoidal algorithms at low bit rates by embedding additional signal models in the coder architecture. The motivation for this work is as follows. Whereas the waveform-preserving perceptual transform (Chapter 7) and subband (Chapter 8) coders tend to target transparent quality at bitrates between 32 and 128 kb/s per channel, the sinusoidal coders proposed thus far in the literature have concentrated on low rate applications between 2 and 16 kb/s. Rather than transparent quality, these algorithms have emphasized source robustness, i.e., the ability to deal with general audio at low rates without constraining source model dependence. Low rate sinusoidal algorithms (ASAC, HILN, etc.) represent the perceptually significant portions of the magnitude spectrum from the original signal without explicitly treating the phase spectrum. As a result, perceptually transparent coding is typically not achieved with these algorithms. It is generally agreed that different classes of state-of-the-art coding techniques perform most efficiently in terms of output quality achieved for a given bit rate. In particular, CELP speech algorithms offer the best performance for clean speech below 16 kb/s, parametric sinusoidal techniques perform best for general audio between 16 and 32 kb/s, and so-called “time-frequency” audio codecs tend to ...