Understanding the quality losses of embedded speech and audio coders

At a given bitrate, and for a given coding structure, an embedded speech coder often shows a loss in quality when compared to a corresponding non-embedded coder. This paper presents a general framework to investigate how this loss depends on common design choices such as the frame length, prediction order and quantizer dimensions used by the coders. The results show that predictive and transform predictive coding schemes, and schemes which normalize energy over short intervals in time and frequency, may see smaller losses when used in embedded designs than compared to other coding paradigms.