#8 in a series of articles about wander and jitter
Although I am guessing, I don’t think that it is crazy to say that the majority of digital audio systems today employ some kind of sampling rate conversion somewhere in the signal flow.
A sampling rate converter is a physical device or a processing block in some software that takes an audio signal that has been sampled at one rate (say, 44.1 kHz) and converts it to an audio signal at another rate (say, 48 kHz).
There are many reasons why you might want to do this. For example, if you have a device that has equalisation (filtering), then if you change the sampling rate, you will have to new coefficients into the filters. If you have a LOT of filters, then it might take so much time to load them into the system that you’ll miss the first second or two of a song if it’s a different sampling rate than the previous song. So, instead of doing this, you keep your processing at one constant (or ‘fixed’) sampling rate, and convert the input to that rate. This might even be true in the case where the incoming sampling rate is the same as the internal sampling rate. For example, you might be “sample rate converting” from 48 kHz to 48 kHz – just to keep the design of the system clocking constant.
Looking very broadly, there are two options for sampling rate conversion.
Synchronous Sampling Rate Conversion
Let’s say that you have to convert from 48 kHz to 96 kHz – a multiplication of 2. In this simple case, you could take the incoming samples, and insert an new, extra one mid-way between each of them. The value of the new sample depends on how you are doing the math to calculate it. We will not discuss this here. The important thing about this concept is that the timing of the output is “locked” to the input. In this example, every second sample of the output happens at exactly the same time as every sample at the input. This can also be true if the ratio of the sampling rates are not “nicely” related like a 2:1 ratio. For example, if you have an input at 44.1 kHz and and output at 48 kHz, you could take the incoming 44.1 kHz signal, insert 47999 “virtual” samples between each of the original samples (making the new sampling rate 2116800000 Hz) and then pull an output sample from that stream every 444100 samples.
In other words:
(44100 * 48000) / 44100 = 48000
Of course, this is not a smart way to do this (because it will be a huge waste of processing power and memory – and imagine how big the numbers would be if you’re converting 176.4 kHz to 192 kHz… bigger!), but it would work, as long as the “virtual” samples you create at the very high “virtual” sampling rate have the correct values.
This type of sampling rate conversion, where the output is numerically “locked” to the input in time (meaning that, at some regular interval of time, the input and the output samples will happen simultaneously – or at least with a constant delay) is called synchronous sampling rate conversion. It’s called that because the input and the output are synchronised with each other… A bit like gears meshing together.
Asynchronous Sampling Rate Conversion
There is another way to do this, where we do not lock the output clock to the input clock. Let’s say that you want to build a device that has a constant sampling rate at its output, but you don’t really know what the sampling rate of the input is. In this case you will use an asynchronous sampling rate converter – so-called because there is no fixed lock between the input and output clocks.
In this case, the incoming signal is analysed and its sampling rate is measured. The way this is done is a little similar to the method shown above. You take the clock running at the rate of the output’s signal and multiply that by some value (say 512, for example) to create an internal “virtual” clock running at a higher sampling rate. You then “grab” the value of an incoming sample and apply its value to the “virtual” sample that is closest in time. This allows the incoming samples to drift in time relative to the output samples.
In both cases, there is the open question of how you generate the signal at the higher internal sampling rate. This can be done using a kind of low pass filter that is effectively similar to the reconstruction filter in a DAC. I will not talk about this any more than that – other than to say that the response characteristics of that filter are VERY important… So, if you’re planning on building your own sampling rate converter, read a lot more stuff on the subject than what I’ve written here – because what I’ve written here is most certainly not enough information.
There’s one strange effect that pops up here. Since, in an ASRC (Asynchronous Sampling Rate Converter) the incoming signal is sampled at discrete times that are numerically related to the output sampling rate, then any potential jitter in the system is also quantised in time. So, for example, if your output sampling rate is 48000 samples per second, and you’re creating the internal sampling rate by multiplying that by 512, then any jitter in the ASRC cannot have a value less than 1/(48000*512) second = 4.069*10^-8 or 40.69 nanoseconds. In other words, in such a system, the error caused by jitter will be 0, ±40.69 nanoseconds, ±81.38 nanoseconds, and so on. It can’t be something in between… (assuming that the output clock is perfect. If it’s drifting due to jitter, then those values will also drift…)
The good news is that, if the clock that is used for ASRC’s output sampling rate is very accurate and stable, and if the filtering that is applied to the incoming signal is well-done, then an ASRC can behave very, very well – and there are lots of examples of this. (Sadly, there are many more examples where an ASRC is implemented poorly. This is why many people think that sampling rate converters are bad – because most sampling rate converters are bad.) in fact, a correctly-made sampling rate converter can be used to reduce jitter in a system (so you would even want to use it in cases where the incoming sampling rate and outgoing sampling rates are the same). This is why some DAC’s include an ASRC at the input – to reduce jitter originating at the signal source.
Wrapping up Part 8: The take-home messages for these three parts in Section 8 are:
- Sampling Jitter results in some kind of distortion of the signal that can be related to the signal itself
- Sampling Jitter can occur in the ADC, the DAC, or an ASRC
- If implemented correctly, an ASRC can be used to attenuate jitter in a system
- Once introduced to the signal, jitter cannot be attenuated. So, if you have a recording that was made using an ADC with a lot of jitter, the artefacts caused by that jitter is in the recorded signal forever. If you have a DAC that has absolutely no jitter whatsoever (this is not possible) then this will not eliminate the jitter that is already in the signal. Of course, it won’t make the situation worse… but it won’t make it better.
Addendum. If you want to dig further into the world of Sampling Jitter and the advantages of using ASRC’s to attenuate jitter, I highly recommend the following as a good starting point:
- Julian Dunn’s paper called “Jitter Theory” – Technical Note TN-23 from Audio Precision. This is a chapter in his book called “Measurement Techniques for Digital Audio”, published by Audio Precision. See this link for more info.
- Clock Jitter, D/A Converters, and Sample-Rate Conversion
By Robert W. Adams, Published in The Audio Critic, Issue No. 21
- The Effects of Sampling Clock Jitter on Nyquist Sampling Analog-to-Digital Converters and on Oversampling Delta Sigma ADCs, Steven Harris. AES Preprint #2844 (87th International Convention of the AES, October 1989)
- Jitter Analysis of Asynchronous Sample-rate Conversion, Robert Adams. AES Preprint #3712 (95th International Convention of the AES, October 1993)