Perceptual Evaluation of the Effects of Dither on Low Bit Rate PCM Systems

When a signal, such as a speech waveform, is quantized, the quantization error waveform is usually correlated with the original signal. This correlation is virtually inperceptible when the quantization is quite fine-i.e., a large number of bits/sample. For crude quantizations, however, the correlation becomes quite large and the quantization error is easily perceived. As a result, it can become quite disturbing to listen to speech quantized to a low number of bits/sample for an extended period of time. In such cases, techniques that decorrelate the quantization error from the signal are attractive, even if they do not increase the signal-to-noise ratio of the system. Dithering is such a technique in 1487 1488 T H E BELL SYSTEM TECHNICAL J O U R N A L , SEPTEMBER 1972 which a pseudo-random "dither" noise is added to the speech before quantizing, and then the identical noise is subtracted producing a quantization error which is uncorrelated with the original speech waveform. 1 Figure 1 shows a comparison between straight P C M and a system in which dithering is used. In an earlier work, Jayant and Rabiner 2 discussed several theoretical issues involved with dithering and demonstrated its utility for the quantization of speech signals. In this paper, we present experimental results on the perceptual effects of dither on both the preference and intelligibility of P C M encoded speech.