Fig.7.1 illustrates the block diagram of a bitstream D/A conversion system[12]. This system consists of a digital filter, a noise shaper and a reconstruction

filter. Eventually, the power amplifier can be integrated on the same chip. In the digital filter the data rate of the input digital word is increased by interpolative upsampling 64 times. The filter interpolates the input signal and calculates intermediate points. The low-pass filter limits sharply the audio-band at 20KHz. In order to reduce the quantization noise in the baseband two techniques are used: oversampling and noise shaping.

By oversampling in the digital filter, the quantization is performed at higher clock rates and the resulting quantization noise power is spread over a larger bandwidth. This yields lower noise in the baseband. Noise shaping is a technique used to reduce the quantization noise in the audioband, by shaping the quantization noise out of the baseband. In combination with oversampling this method gives sufficient reduction of the quantization noise to realize high accuracy systems. In the 1-bit D/A converter, the digital sequence present at the output of the noise-shaper (1,0,0,1,0…) it is translated into an accurate two-level analog signal (A,-A,-A,A,-A,…) with high linearity. In the same block, a sampled data FIR filter (LPD) will suppress the out-of-band noise. To reduce harmonic distortion and intermodulation products in the output power amplifier, the level of the high frequency quantization noise has to be lower than -50dB. A first order, continuous-time analog low pass filter (LPA) will reconstruct the signal by attenuating the spectral repetitions at multiples of sampling frequency.

In fig.7.1 the signal and the noise spectra of the D/A system are shown. For simplicity, the oversampling frequency is not drawn at the right scale (4 times instead of 64 times). Fig.a shows that a sampled signal consists of an infinite sequence of the original spectra shifted by multiples of f_s. Therefore also the noise power is shown up to ½f_s. Oversampling (see fig.b) spreads the noise over a larger frequency band and places the signal spectra further from each other. The noise shaper shapes the noise out of the baseband to higher frequencies (see fig.c). In fig.d+e, the signal at the output of the FIR filter (doted) and the signal at the output of the LPA filter (thick lines). The noise at high frequencies is filtered by the time discrete filter. The final operation, i.e. the low-pass continuous-time filtering, removes the undesired spectral repetitions from the signal.