I have AM demodulated voice samples that come from a 14bit ADC that I will filter, process, and then send to a CODEC. I can choose a sample rate of 96kSPS, 64kSPS, 48kSPS, or 32kSPS at the CODEC. Taking into account most of the voice spectrum is limited to less than 15kHz, would I gain anything by sampling at 96kSPS vs 32kSPS? I will use either an IIR or an FIR filter between the ADC and CODEC.

For a lower sample rate, the digital filtering would require more taps to acheive the same performance; however, there would be less samples to filter so that trade off seems like a wash.

  • 1
    $\begingroup$ Are you an audiophile? $\endgroup$
    – user28715
    Jun 28, 2017 at 2:47
  • $\begingroup$ for the same wordsize, higher sampling rate increases the SNR (within a fixed bandwidth) by 6 dB or 1 added bit of resolution for every two octaves of sample rate. and that's without noise-shaping. toss in noise shaping and you get better than 3 dB increase in SNR for each doubling of sample rate. $\endgroup$ Jun 28, 2017 at 3:29
  • $\begingroup$ Not an audiophile, but I do need to understand the tradeoff. $\endgroup$
    – jeremy
    Jun 28, 2017 at 17:48

1 Answer 1


Sampling at a higher rate will distribute the quantization noise over a wider frequency, thus reducing the noise spectral density due to quantization noise specifically, with a lot of caveats. For more details on that see What are advantages of having higher sampling rate of a signal?

So when the SNR is limited by quantization noise, then increasing the sampling rate can increase the SNR by reducing that portion of quantization noise that is in the signal bandwidth. However if the quantization noise is significantly less (such that we are actually sampling the actual in-band noise), then increasing the sampling rate will not improve that in-band SNR.

In your last paragraph, if you are referring to running the same filter at a lower rate versus a higher rate, at a lower sampling rate the digital filter will require LESS taps for the same filter performance. For more details on that see Filter Order Rule of Thumb and How many taps does an FIR filter need?.

An effective strategy to consider is to use a higher sampling rate which relaxes the analog filtering requirements and can give you more effective bits if needed at the analog to digital boundary (which are realized after subsequent filtering digitally). This is then followed by efficient resampling techniques in the digital domain to get to a lower sampling frequency prior to providing the final filtering as required in your system design (coarse and simpler filters are done at the higher rate, and then final "shaping filters" are done at the lowest rate possible thus minimizing number of taps, power dissipation and resources).

For a further details on decimating to a lower rate for improved filter performance see Fast Integer 8 Hz 2nd Order LP for Microcontroller.

  • $\begingroup$ At this point I believe we have to underline the discrimination that what is improved by oversampling quantizers is actually the SQNR (signal to quantization noise ratio) and not the SNR related to the inband analog noise present prior to sampling. Whether oversampling can also be utilised to improve that more fundamental SNR is another concern. Your answer clearly tells it but the question title creates a wrong implication I think. $\endgroup$
    – Fat32
    Jun 28, 2017 at 19:12
  • $\begingroup$ @Fat32 This is an excellent comment! Also after rereading the question I was confused as to where the sample rate is occurring, as it almost sounds like he is sampling at some unspecified rate, then resampling to the CODEC output rate with filtering in between (in which case his comment about filter taps would be more applicable). So I am not confident now that I completely answered his question. Related to Fat32's comment, I always choose the quantization noise floor (in band) to be 10 to 15 dB below my signal noise floor. A 10 dB margin for example raises the noise 0.4 dB. $\endgroup$ Jun 29, 2017 at 14:47
  • $\begingroup$ In fact I'm even more confused now. @jeremy should make his question more clear. Where is AM demodulation taking place? Analog or digital? Hence at which side the ADC resides? before (digital) demodulation or after (analog) demodulation? What is this CODEC all about, a DAC? What he means by selecting sampling rate at the CODEC? Anyway as he's already confirmed your answer as correct and relevant, therefore I think what you have assumed was actually what he was aiming to ask :-) $\endgroup$
    – Fat32
    Jun 29, 2017 at 15:03
  • $\begingroup$ @Fat32 He says it is AM demodulated voice samples from an ADC, so the digital samples are baseband audio, that much is clear. Then he says choosing the sampling rate at the CODEC so not clear that that has to be the same rate as the ADC, my answer was specific to the ADC rate. (although likely applicable as well to the CODEC rate). But I am not knowledgeable about everything that goes on in the CODEC at the different rates that could also effect the output SNR of the audio. There may be an opportunity for a better answer here if that is really his question. $\endgroup$ Jun 29, 2017 at 15:11
  • 1
    $\begingroup$ Yes that's the most reasonable assumption, I would also make. But I still believe he must make it clear at which stage he's looking for an SNR improvement and what type of an SNR is that he wants to improve. He should better put an informative block diagram of the overall signal processing chain and clearly describe the data formats that's entering and leaving those blocks. Nevertheless, he seems to be happy with the provided answer and therefore I deduce he was talking about initial ADC sampling rate and hence SQNR improvement at the input stage is what he would get from oversampling there. $\endgroup$
    – Fat32
    Jun 29, 2017 at 15:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.