Why ADC converters do not use proper anti-alias filters?
As we all know, the Nyquist theorem states a necessary condition of proper sampling - input signal must not contain frequencies above Fs/2, half of sampling frequency. If this condition is violated, we are experiencing alias frequencies that are not present in the input signal and the spectrum of the sampled signal is thus contaminated. The worst situation is when input signal frequency, Fin and sampling frequency Fs are related so that
Fin - N*Fs < Fs/2
This is especially critical when we measure class D amplifiers, as they always have considerable level of switching frequency residuals in their output spectrum. I will demonstrate this issue on my PMA-NC252MP amplifier measured through E1DA Cosmos ADC designed by @IVX .
I measured switching frequency residuals at NC252MP output, it was about 350 mVrms with frequency around 380 - 400 kHz, and this frequency fluctuates with output signal level.
Let's relate it to N*Fs, for sampling frequency 48kHz and N = 8 we get 384kHz and alas, 384kHz subtracted from switching frequency falls to the audio band. With respect to the fact that the switching frequency floats with signal level, the alias frequency will travel through the audio band. For 96kHz or 192kHz sampling frequency, we have similar issues.
Enough talking and let's see some measurements.
As a signal source, I have used my analog generator which operates in range of frequencies from 10Hz to 1MHz. It has distortion not better than -90dB, but for this test it is more than good enough. Output voltage level was set to 377mVrms, similar as level of switching residual from NC252MP. The generator frequency was set to ~400kHz, similar as NC252MP switching frequency. 4KHz signal of same level 377mV was used as well as a reference. ADC used was E1DA Cosmos ADC, set to 2.7V input range and provided with an additional input resistor divider that makes the ADC full scale range 0dBFS = 45.547V. As a comparison, the same measurement configuration was used, but with LC low pass filter added, which has flat response to 20kHz and then the slope -40dB/decade. This filter should attenuate 400kHz of some -46dB. The filter was described here in another thread:
www.audiosciencereview.com
Measurements
4kHz / 377 mV sine as a reference
Signal level is -41.6 dBFS
400kHz / 377 mV sine (without LC low pass filter)
We can see alias frequency near 10kHz with amplitude of -98.4 dBFS. So the alias is suppressed by the ADC alone of 56.8dB.
400kHz / 377 mV sine (with the LC low pass filter)
Now the alias amplitude is -140.1 dBFS. Additional suppression with the low pass filter is 41.7dB.
Now let's see the comparison in NC252MP intermodulation distortion measurement, 60Hz + 7kHz SMPTE IMD, 77W/4ohm. We can see the major contribution of the LC low pass filter that lies in the higher dynamic range of the measurement and noise floor decreased of about 12 dB.
Conclusion
Though SOTA DAC converters are usually equipped with good reconstruction filters that effectively suppress the mirror images, the same cannot be said about SOTA audio ADC converters. By definition, sampled signal must be frequency limited to <Fs/2, however it does not happen effectively enough and ADCs usually have no input filters or simple filters, which does not prevent them from creation of alias frequencies that fall into the audio band and into converters usable dynamic range. I understand that it is not easy to provide the ADC with switchable input filters, however designers should consider this request that reflects definition of proper sampling. This is especially an issue when measuring class D amplifiers or components with noise shaped output. The situation becomes worse and worse as the level of the useful input audio signal is decreasing at the constant level of the switching residuals.
As we all know, the Nyquist theorem states a necessary condition of proper sampling - input signal must not contain frequencies above Fs/2, half of sampling frequency. If this condition is violated, we are experiencing alias frequencies that are not present in the input signal and the spectrum of the sampled signal is thus contaminated. The worst situation is when input signal frequency, Fin and sampling frequency Fs are related so that
Fin - N*Fs < Fs/2
This is especially critical when we measure class D amplifiers, as they always have considerable level of switching frequency residuals in their output spectrum. I will demonstrate this issue on my PMA-NC252MP amplifier measured through E1DA Cosmos ADC designed by @IVX .
I measured switching frequency residuals at NC252MP output, it was about 350 mVrms with frequency around 380 - 400 kHz, and this frequency fluctuates with output signal level.
Let's relate it to N*Fs, for sampling frequency 48kHz and N = 8 we get 384kHz and alas, 384kHz subtracted from switching frequency falls to the audio band. With respect to the fact that the switching frequency floats with signal level, the alias frequency will travel through the audio band. For 96kHz or 192kHz sampling frequency, we have similar issues.
Enough talking and let's see some measurements.
As a signal source, I have used my analog generator which operates in range of frequencies from 10Hz to 1MHz. It has distortion not better than -90dB, but for this test it is more than good enough. Output voltage level was set to 377mVrms, similar as level of switching residual from NC252MP. The generator frequency was set to ~400kHz, similar as NC252MP switching frequency. 4KHz signal of same level 377mV was used as well as a reference. ADC used was E1DA Cosmos ADC, set to 2.7V input range and provided with an additional input resistor divider that makes the ADC full scale range 0dBFS = 45.547V. As a comparison, the same measurement configuration was used, but with LC low pass filter added, which has flat response to 20kHz and then the slope -40dB/decade. This filter should attenuate 400kHz of some -46dB. The filter was described here in another thread:
What is on your workbench right now?
The lab might be tidy but the bedroom bench certainly isn't. :rolleyes: A little bit of everything, really. These are an ADAU1467 board, ADAU1701 and an Amanero USB-I2S interface clone.
www.audiosciencereview.com
Measurements
4kHz / 377 mV sine as a reference
Signal level is -41.6 dBFS
400kHz / 377 mV sine (without LC low pass filter)
We can see alias frequency near 10kHz with amplitude of -98.4 dBFS. So the alias is suppressed by the ADC alone of 56.8dB.
400kHz / 377 mV sine (with the LC low pass filter)
Now the alias amplitude is -140.1 dBFS. Additional suppression with the low pass filter is 41.7dB.
Now let's see the comparison in NC252MP intermodulation distortion measurement, 60Hz + 7kHz SMPTE IMD, 77W/4ohm. We can see the major contribution of the LC low pass filter that lies in the higher dynamic range of the measurement and noise floor decreased of about 12 dB.
Conclusion
Though SOTA DAC converters are usually equipped with good reconstruction filters that effectively suppress the mirror images, the same cannot be said about SOTA audio ADC converters. By definition, sampled signal must be frequency limited to <Fs/2, however it does not happen effectively enough and ADCs usually have no input filters or simple filters, which does not prevent them from creation of alias frequencies that fall into the audio band and into converters usable dynamic range. I understand that it is not easy to provide the ADC with switchable input filters, however designers should consider this request that reflects definition of proper sampling. This is especially an issue when measuring class D amplifiers or components with noise shaped output. The situation becomes worse and worse as the level of the useful input audio signal is decreasing at the constant level of the switching residuals.
