On Floaking-Point Sigma-Delta D/A Conversion

Digital audio has rapidly replaced andog audio ining the demand for high-quality analog-tcwdigital (A/D) and digital-to-analog (D/A) converters. The pmformance of conventional Nyquist rate PCM converters is limited by the need of a high-order analog anti-aliasing filter, a reconstruction filter and a sample-and-hold amplifier. These limitations CM be eliminated and better performance can be achieved by using oversampled sigmadelta (SA) converters which convert the signal to a high-frequency one-bit streem. S A converters utilize a noise shaping feature in which quantization errors are shaped to high fiequencies and removed with a digital lowpass filter [1]-131. Fkently, oversampled CA converters have emerged in numerous digital audio equipments, such as CD and DAT players. Currently high quality audio converters are based on uniform, fixed-point quantization scheme, its the dynamic range and the signal-to-noise ratio (SNR) depend on the conversion accuracy. The SNX depends also on the signal level; hence it decreases with low signal levels. Unifonn quantization is suitable for signals which distribute evenly in the converter amplitude range. Music signals do not however have a uniform distribution and therefore new quantization schemes are desired if a better SNR behavior is wanted. Better performance can be obtained if the quantization level are spread approximately logarithmically. This can be irchieved by using compending converters or floating-point converters. Examples of non-uniform quantizers are found in PCM telephones, and NICAM [4]-[5] and DIGICIPHERTM [SI television sound systems. In recent years, the use of floating-point arithmetic in digital signal processing (DSP) has rapidly i n c r e a d due to the development of fast and low-cost floating-point signal processors. Since the algorithms are based on floating-point arithmetic, a need for floating-point converters has arisen. Eventually even such functions as volume control are likely to be implemented in digital form requiring a larger dynamic range for the D/A converter than is today commonly used. Even there are some difficulties in implementing Nyquist rate floating-point converters, they have beein studied extensively [A, [SI. In this paper, we discuss the design and implementation aspects of oversempled C A D/A conversion based on a full floating-point number system.