This example shows how to model analog-to-digital conversion using a sigma-delta algorithm implementation.
The Oversampled Sigma-Delta A/D Converter is a noise-shaping quantizer. The main purpose of noise-shaping is to reshape the spectrum of quantization noise so that most of the noise is filtered out of the relevant frequency band, for example, the audio band for speech applications. The main objective is to trade bits for samples; that is, to increase the sampling rate but reduce the number of bits per sample. The resulting increase in quantization noise is compensated by a noise-shaping quantizer. This quantizer pushes the added quantization noise out of the relevant frequency band and thereby preserves a desired level of signal quality. This reduction in the number of bits simplifies the structure of A/D and D/A converters.
As seen in this example, the analog input is prefiltered by an antialiasing prefilter whose structure is simplified because of oversampling. The input signal is oversampled by a factor of 64. The Integrator, 1-Bit Quantizer, and Zero-Order Hold blocks comprise a two-level analog to digital converter (ADC). The output of the Zero-Order Hold is then subtracted from the analog input. The feedback, or approximation, loop causes the quantization noise generated by the ADC to be highpass filtered, pushing its energy towards the higher frequencies (
64*fs/2) and away from the relevant signal band. The decimation stage reduces the sampling rate back to 8 KHz. During this process, it removes the high frequency quantization noise that was introduced by the feedback loop and removes any undesired frequency components beyond
fs/2 (4 KHz) that were not removed by the simple analog prefilter.
The example versions illustrate two possible decimator design solutions.
The floating-point version model uses a cascade of three polyphase FIR decimators. This approach reduces computation and memory requirements as compared to a single decimator by using lower-order filters. Each decimator stage reduces the sampling rate by a factor of four. The latency introduced by the filters is used to set the appropriate 'Time Delay' in the 'Transport Delay' block. The three FIR Decimation filters each introduce a latency of 16 samples, due to the group delay of the filter (the actual value of 15.5 is rounded up to the nearest integer number of samples). Due to the decimation operation the total latency introduced by the three filters is as follows: 16 (first filter) + 4*16 (second filter) + 16*16 (third filter) to give a final total delay of 336. The denominator of the 'Time delay' parameter is the base rate of the model (512 kHz).
The fixed-point version uses a five-section CIC decimator to reduce the sampling rate by the same factor of 64. While not as flexible as a FIR decimator, the CIC decimator has the advantage of not requiring any multiply operations. It is implemented using only additions, subtractions, and delays. Therefore, it is a good choice for a hardware implementation where computational resources are limited. The CIC Decimator introduces a latency of 158 samples, which is the group delay of the filter (157.5) rounded up to the nearest integer. This is the value used in 'Time Delay' parameter of the 'Multistage CIC Processing Delay' block.
Orfanidis, S. J. Introduction To Signal Processing, Prentice Hall, 1996.
Floating-point version: dspsdadc
Fixed-point version: dspsdadc_fixpt