I’m originally from Newfoundland – one of the few places in the world with a 1/2-hour time zone. So, when it’s 10:00 a.m. in Montreal, it’s 11:30 a.m. in St. John’s – my home town. This meant that, when I was a kid 40 years ago, and we would call our relatives in Toronto or Germany to wish them a Merry Christmas, there were two questions that you could always rely on being asked: (1) what’s the weather like there? and (2) what time is it there?
These days, I have a similar problem that is well-described by “Segal’s Law“. My iPhone and my wristwatch (an old analogue one with hands that go around pointing at the floor and the fridge…) are never synchronised… This is because of two things: (1) I probably did a bad job of setting my watch and (more importantly) (2) my watch runs just a little bit slowly…
So, let’s say, for example, that I set my watch to be EXACTLY in sync with my phone on a Monday morning at 9:00 a.m. As the week goes by, my iPhone and my watch drift apart, and, just for the sake of argument let’s say that, one week later, when my iPhone turns over to 9:00 a.m. on Monday morning, my wristwatch turns over to 8:59 a.m. So, I lose 1 minute per week on my watch.
(It’s pretty safe to assume that my iPhone is also not perfect – but it’s different because, every once in a while, it compares its internal clock with another, more accurate clock somewhere else via a connection across the Internet (which, we will assume, for the purposes of this discussion, works).)
Let’s consider this from a strange point of view. Let’s assume that
- I’m checking the time on my watch every minute, on the minute
- someone else is “fixing” my watch every week so that it’s correct at 9:00 a.m. on Mondays. They do this by adjusting the watch to the correct time 30 seconds before the iPhone says it’s 9:00 a.m.
- I don’t know that they’re doing this for me…
If we think about this from my perspective, I’ll live in a strange world where 8:59 on Mondays never exists. This is because at 8:58 and 30 seconds (on my watch), my friend re-sets the time to 8:59 and 30 seconds (while I’m not looking) to synchronise with the iPhone…
IF my watch was running fast – say, gaining one minute each week, then I would live in a different strange universe where 9:00 happens twice every Monday morning…
The basic problem here is that we have two clocks that do not run at the same rate – but they are expected to do so. So, we synchronise them regularly (in the above example, on Monday mornings at 9:00) – but between those synchronisation events, they drift apart in time.
The example above is very, very similar to the way a digital audio streaming system works – especially if you’re using a wireless connection between the transmitting device and a receiver.
Lets say that you’re playing a sound file that was recorded at 44.1 kHz and streaming it wirelessly to a receiver. I’m trying to be as generic as possible here, but I could be talking about a Bluetooth connection to a pair of headphones or a WiFi connection via DLNA to a device connected to a pair of loudspeakers, for example…
It is not unusual with such a connection for the transmitter to collect up a block of audio samples – say, 64 of them – and send them to the receiver’s input buffer. The receiver then pulls those samples out, one by one, and (eventually) sends them to a digital-to-analogue converter that produces a signal that (eventually) comes out as an audio signal. Then, 64/44100’ths of a second later (64 samples later) the transmitter sends another block, and so on and so on until the song ends.
This system works well if the clock inside the transmitter and the clock inside the receiver are perfectly synchronised. We can even be a little generous and say that they can drift apart a little – but not so much that we either run out of samples to play (because the receiver is playing them out faster than they’re coming in from the transmitter) or that we have samples left over to play when the next block comes in (because the receiver is playing them out slower than they’re coming in from the transmitter).
Dealing with this problem the right way
The right way to deal with this issue is for the receiver to always be checking what time it thinks it is when the block arrives from the transmitter. If the block arrives a little early, then the receiver should think “hmmmm, my clock is going too slowly – I’ll speed it up a bit”. If the block arrives a little late, then the receiver should adjust its clock to go a little slower.
So, in this case, the receiver has a basic, nominal speed for its internal clock – but it’s constantly adjusting it to be faster and slower to try and match the clock of the transmitter – but it can only do this adjustment at the block rate – the frequency at which the blocks of samples arrive, which is dependent on the block length (how many samples are in each block) and the sampling rate (how many samples per second). (Of course, this can result in “jitter and wander” problems if you’re not careful (I won’t talk about this here…) – so you have to pay a little attention to how quickly you’re adjusting your clock rate… but that’s “just” a matter of correct implementation.)
Dealing with this problem the wrong way
There is another way to deal with this problem, which, unfortunately, has measurable and possibly audible consequences. This implementation is basically the same as my original example, where I had a friend “fixing” my wristwatch once a week. You have a transmitter that sends blocks of samples to the receiver – and although these two devices should have exactly the same clock rate, they don’t.
Let’s say, for example, that the receiver is playing the samples faster than they’re being sent by the transmitter. This means that the two will slowly drift farther and farther apart until, eventually, the receiver will have to play a sample, but nothing has come in from the transmitter yet, so there’s no sample there to play. In this case, the receiver says “no problem, I’ll just play the last sample again, and the next block will come in while I’m doing that” – so it inserts an extra sample that is just a duplicate of the previous one.
If the receiver’s clock is going slower than the transmitter’s, then, as the two drift farther apart, we will get to a moment where the receiver will receive a new block of samples but it’s not done playing all of the samples in the previous block yet. In this event, it says “no problem, I’ll just leave that last sample out and move on to the next block to catch up” – so it skips a sample.
This is called a “Skip / Insert” strategy for dealing with clock synchronisation. It’s done by software and hardware engineers because it’s simple to implement, and, in many cases, a manufacturer can get away with this, since it is rarely audible for a couple of reasons.
Can this be measured?
The simple answer to this is “yes” – and it can be measured in a number of different ways. I’ll show one way below…
Can I hear it?
The honest answer to this question is “sometimes” – but it’s not as easy to detect as one might think. Of course, a skip/insert event (a duplicated sample or a dropped one) creates an artefact. However, the magnitude of this artefact relative to the “correct” signal is dependent on when it happens.
Let’s take a look at a couple of simple cases. We’ll “transmit” one period of a sine wave that should come out on the other side of the system looking like Figure 1.
But what happens if we don’t get a block in time to keep outputting a signal? We insert a duplicate sample and hope that the block comes in before I have to send out another one. Examples of this are shown in Figures 2 and 3, below.
You’ll probably notice that it’s much easier to see which sample I duplicated in Figure 3 than in Figure 2. In Figure 3 it was sample number 26 that was duplicated. In Figure 2 it’s sample number 13.
The reason it’s easier to see the error in Figure 3 is that duplicating the sample causes an obvious change in the slope of the signal, whereas in Figure 2 it does not – the slope of the signal is 0, and by duplicating a sample, I am also making it 0 – but for a slightly longer time.
This does not mean that we did not generate an error. It just means that we’ll probably “get away with it” in the case of Figure 2, and we probably won’t in the case of Figure 3.
However, since the drifting of the two clocks (in the receiver and transmitter) are not dependent on the signal, there’s no way to know when this is going to happen.
And, of course, if this happens in the middle of a snare drum hit or a ssssinger sssstarting a word in a ssssong with the letter “s” – then we also won’t hear it because there’s so much going on (frequency-wise) that the artefact will be buried in the mess.
Also, since this clock drifting is usually not completely regular, the errors do not usually come in at a regular rate (although I’ve seen exceptions…). So, it’s not like you can listen for “a click every second” or “one per minute”. They happen when they happen – hopefully when you’re not listening and/or when the tune is busy enough to hide it.
A skip event is similar to an insert, as you can see in the two examples in Figures 4 and 5.
Again, I’ve intentionally put in these two skips in places where they are least obvious (Figure 4) and most obvious (Figure 5).
The real world
One of the tests that can be done on an audio system is to send a sinusoidal signal with a swept frequency through a system, capture the output, and then do a spectrogram of the result. In theory, if you see anything other than a single frequency at any one time at the output, then you know that something has happened to the signal. You would probably then need to go back and look at the output signal itself to start evaluating exactly what happened… This is a test that is used to evaluate one aspect of the performance of different sampling rate converters, for example, at this site.
Let’s take a sine sweep and run it through a system. The sweep goes up logarithmically in frequency from 20 Hz to about 90% of Nyquist (which would correspond to 20,000 Hz in a system running at 44.1 kHz) over 60 seconds and has a level of -1 dB FS. We’ll then capture the output in a system that is behaving perfectly and do a spectrogram of this, looking for artefacts down to some level below the signal level. (If you’re really geeky, you’ll know that this signal-to-error ratio is dependent on the window length of the FFT I’m using to create the spectrogram – but this is beyond our discussion today…).
An example of the output of a system that is behaving well is shown in Figure 6.
You may notice that the plot looks a little “wide” in the beginning. This is because the window length of the FFT I’m using to analyse the signal isn’t long enough to get a precise analysis of a low-frequency signal. So, this is an artefact of the analysis – not an error in the playback system.
What happens if we have random skip/insert events in the system? This is shown in Figure 7.
The signal in Figure 7 was one that I created – I intentionally made skip/insert events at random times and applied them to my test signal.
There are two things to notice here. The first is that each event is visible as a vertical “spike” in the plot. This is because a skip/insert event will cause a short, wide-band “burst” that sounds like a click. However, the bandwidth of the click is dependent on when it happens relative to the signal. For example, the skip/insert events in Figure 2 and 4 would not create as much high-frequency energy as the ones in Figure 3 and 5. So, the bigger the effect on the slope of the signal, the more high frequency energy we’ll get in our “click” sound. Since the slope of a signal increases with frequency, then this also means that low-frequency signals will likely produce lower-bandwidth artefacts.
Now let’s look at the results from some real-world devices and systems that are commercially available.
As you can see in Figure 8, there was one skip/insert event that happened during the 60 seconds I was running this test. Remember that the time that that event happened had nothing to do with the frequency it was playing. It just happens when it happens due to the relationship between the transmitter’s and the receiver’s clock speeds.
Figure 9 shows the results from a different system/device that obviously uses a skip/insert strategy to deal with clock synchronisation problems. It also obviously has some serious clock issues, since it has to correct on the order of approximately once a second…
Figure 10 shows the results from a different system/device that uses a skip/insert strategy – but appears to do so at scheduled intervals. In this case, there is a high probability of getting a skip/insert event every 10 seconds with the counter starting at the instant I starting hearing the music.
Inquisitive readers may be asking why it is that, although I’m doing an analysis down to -101 dB FS (100 dB below the signal level of -1 dB FS), you can’t see the effects of the dither noise floor in my original 16-bit file (which is normally assumed to be at -93 dB FS). This is because the -93 dB FS estimate of a dither signal assumes that you are looking at the total energy from the entire frequency band. The spectrograms above are based on FFT’s that split up the total frequency band into “slices” (called frequency bins) – and the total energy in each of these bins is less than the total energy in all of them (one person clapping is not as loud as 1000 people clapping at the same time…). If we wanted to see the dither noise, I would have had to set my analysis to go down approximately 30 dB lower – but the actual value for this is dependent on the relationship between the sampling rate, the window length of the FFT’s, and the windowing function that I’m using.
Do not bother contacting me to ask which “commercially-available system/device” I measured and in which I found these errors. I’m not doing this to get anyone in trouble. I’m just doing this to try to illustrate common errors that I see often when I evaluate and test audio devices.
An besides, it would not be fair for me to rat on specific companies, systems, or devices, since, in some cases, these errors may have already been fixed with a firmware update, meaning that “naming names” would be irrelevant and unnecessarily detrimental.
But, I will say that I see this problem often. A rough estimate is that I would see errors like this on roughly half of the commercially-available devices and systems I test. It can also be sneaky, as we saw in Figures 8 and 10. Sometimes you get one of these clicks only once in a minute. So, if you do a 10-second measurement to test if your wireless audio receiver is “bit accurate” – the answer can be “yes” – but if you keep measuring for 1 or 2 minutes, you find out the answer is “no”…
If it helps, I could have used the example of a leap year instead of two clocks at the beginning. The reason we have a February 29 every 4 years is that our calendar “runs” a little faster than the time it takes us to get around the sun (because a “year” is actually 365.25 days long…). So, every 4 years we have to “insert” a day to put the two clocks back in sync.
Also, since a “year” is not exactly 365.25 days long, we also have the occasional “leap second” as well. But most people don’t notice this, since it’s rarely useful as an excuse when you’ve missed a meeting…
Reminder: This is still just the lead-up to the real topic of this series. However, we have to get some basics out of the way first…
In the first posting in this series, I talked about digital audio (more accurately, Linear Pulse Code Modulation or LPCM digital audio) is basically just a string of stored measurements of the electrical voltage that is analogous to the audio signal, which is a change in pressure over time… In the second posting in the series, we looked at a “trick” for dealing with the issue of quantisation (the fact that we have a limited resolution for measuring the amplitude of the audio signal). This trick is to add dither (a fancy word for “noise”) to the signal before we quantise it in order to randomise the error and turn it into noise instead of distortion.
In this posting, we’ll look at some of the problems incurred by the way we carve up time into discrete moments when we grab those samples.
Let’s make a wheel that has one spoke. We’ll rotate it at some speed, and make a film of it turning. We can define the rotational speed in RPM – rotations per minute, but this is not very useful. In this case, what’s more useful is to measure the wheel rotation speed in degrees per frame of the film.
Take a look at the left-most column in Figure 1. This shows the wheel rotating 45º each frame. If we play back these frames, the wheel will look like it’s rotating 45º per frame. So, the playback of the wheel rotating looks the same as it does in real life.
This is more or less the same for the next two columns, showing rotational speeds of 90º and 135º per frame.
However, things change dramatically when we look at the next column – the wheel rotating at 180º per frame. Think about what this would look like if we played this movie (assuming that the frame rate is pretty fast – fast enough that we don’t see things blinking…) Instead of seeing a rotating wheel with only one spoke, we would see a wheel that’s not rotating – and with two spokes.
This is important, so let’s think about this some more. This means that, because we are cutting time into discrete moments (each frame is a “slice” of time) and at a regular rate (I’m assuming here that the frame rate of the film does not vary), then the movement of the wheel is recorded (since our 1 spoke turns into 2) but the direction of movement does not. (We don’t know whether the wheel is rotating clockwise or counter-clockwise. Both directions of rotation would result in the same film…)
Now, let’s move over one more column – where the wheel is rotating at 225º per frame. In this case, if we look at the film, it appears that the wheel is back to having only one spoke again – but it will appear to be rotating backwards at a rate of 135º per frame. So, although the wheel is rotating clockwise, the film shows it rotating counter-clockwise at a different (slower) speed. This is an effect that you’ve probably seen many times in films and on TV. What may come as a surprise is that this never happens in “real life” unless you’re in a place where the lights are flickering at a constant rate (as in the case of fluorescent or some LED lights, for example).
Again, we have to consider the fact that if the wheel actually were rotating counter-clockwise at 135º per frame, we would get exactly the same thing on the frames of the film as when the wheel if rotating clockwise at 225º per frame. These two events in real life will result in identical photos in the film. This is important – so if it didn’t make sense, read it again.
This means that, if all you know is what’s on the film, you cannot determine whether the wheel was going clockwise at 225º per frame, or counter-clockwise at 135º per frame. Both of these conclusions are valid interpretations of the “data” (the film). (Of course, there are more – the wheel could have rotated clockwise by 360º+225º = 585º or counter-clockwise by 360º+135º = 495º, for example…)
Since these two interpretations of reality are equally valid, we call the one we know is wrong an alias of the correct answer. If I say “The Big Apple”, most people will know that this is the same as saying “New York City” – it’s an alias that can be interpreted to mean the same thing.
Wheels and Slinkies
We people in audio commit many sins. One of them is that, every time we draw a plot of anything called “audio” we start out by drawing a sine wave. (A similar sin is committed by musicians who, at the first opportunity to play a grand piano, will play a middle-C, as if there were other notes in the world.) The question is: what, exactly, is a sine wave?
Get a Slinky – or if you don’t want to spend money on a brand name, get a spring. Look at it from one end, and you’ll see that it’s a circle, as can be (sort of) seen in Figure 2.
Since this is a circle, we can put marks on the Slinky at various amounts of rotation, as in Figure 3.
Of course, I could have put the 0º marl anywhere. I could have also rotated counter-clockwise instead of clockwise. But since both of these are arbitrary choices, I’m not going to debate either one.
Now, let’s rotate the Slinky so that we’re looking at from the side. We’ll stretch it out a little too…
Let’s do that some more…
When you do this, and you look at the Slinky directly from one side, you are able to see the vertical change of the spring from the centre as a result of the change in rotation. For example, we can see in Figure 6 that, if you mark the 45º rotation point in this view, the distance from the centre of the spring is 71% of the maximum height of the spring (at 90º).
So what? Well, basically, the “punch line” here is that a sine wave is actually a “side view” of a rotation. So, Figure 7, shows a measurement – a capture – of the amplitude of the signal every 45º.
Since we can now think of a sine wave as a rotation of a circle viewed from the side, it should be just a small leap to see that Figure 7 and the left-most column of Figure 1 are basically identical.
Let’s make audio equivalents of the different columns in Figure 1.
Figure 10 is an important one. Notice that we have a case here where there are exactly 2 samples per period of the cosine wave. This means that our sampling frequency (the number of samples we make per second) is exactly one-half of the frequency of the signal. If the signal gets any higher in frequency than this, then we will be making fewer than 2 samples per period. And, as we saw in Figure 1, this is where things start to go haywire.
Figure 11 shows the equivalent audio case to the “225º per frame” column in Figure 1. When we were talking about rotating wheels, we saw that this resulted in a film that looked like the wheel was rotating backwards at the wrong speed. The audio equivalent of this “wrong speed” is “a different frequency” – the alias of the actual frequency. However, we have to remember that both the correct frequency and the alias are valid answers – so, in fact, both frequencies (or, more accurately, all of the frequencies) exist in the signal.
So, we could take Fig 11, look at the samples (the black lollipops) and figure out what other frequency fits these. That’s shown in Figure 12.
Moving up in frequency one more step, we get to the right-hand column in Figure 1, whose equivalent, including the aliased signal, are shown in Figure 13.
Do I need to worry yet?
Hopefully, now, you can see that an LPCM system has a limit with respect to the maximum frequency that it can deal with appropriately. Specifically, the signal that you are trying to capture CANNOT exceed one-half of the sampling rate. So, if you are recording a CD, which has a sampling rate of 44,100 samples per second (or 44.1 kHz) then you CANNOT have any audio signals in that system that are higher than 22,050 Hz.
That limit is commonly known as the “Nyquist frequency“, named after Harry Nyquist – one of the persons who figured out that this limit exists.
In theory, this is always true. So, when someone did the recording destined for the CD, they made sure that the signal went through a low-pass filter that eliminated all signals above the Nyquist frequency.
In practice, however, there are many cases where aliasing occurs in digital audio systems because someone wasn’t paying enough attention to what was happening “under the hood” in the signal processing of an audio device. This will come up later.
Two more details to remember…
There’s an easy way to predict the output of a system that’s suffering from aliasing if your input is sinusoidal (and therefore contains only one frequency). The frequency of the output signal will be the same distance from the Nyquist frequency as the frequency if the input signal. In other words, the Nyquist frequency is like a “mirror” that “reflects” the frequency of the input signal to another frequency below Nyquist.
This can be easily seen in the upper plot of Figure 14. The distance from the Input signal and the Nyquist is the same as the distance between the output signal and the Nyquist.
Also, since that Nyquist frequency acts as a mirror, then the Input and output signal’s frequencies will move in opposite directions (this point will help later).
Usually, frequency-domain plots are done on a logarithmic scale, because this is more intuitive for we humans who hear logarithmically. (For example, we hear two consecutive octaves on a piano as having the same “interval” or “width”. We don’t hear the width of the upper octave as being twice as wide, like a measurement system does. that’s why music notation does not get wider on the top, with a really tall treble clef.) This means that it’s not as obvious that the Nyquist frequency is in the centre of the frequencies of the input signal and its alias below Nyquist.
Reminder: This is still just the lead-up to the real topic of this series. However, we have to get some basics out of the way first…
In the last posting, I talked about digital audio (more accurately, Linear Pulse Code Modulation or LPCM digital audio) is basically just a string of stored measurements of the electrical voltage that is analogous to the audio signal, which is a change in pressure over time…
For now, we’ll say that each measurement is rounded off to the nearest possible “tick” on the ruler that we’re using to measure the voltage. That rounding results in an error. However, (assuming that everything is working correctly) that error can never be bigger than 1/2 of a “step”. Therefore, in order to reduce the amount of error, we need to increase the number of ticks on the ruler.
Now we have to introduce a new word. If we really had a ruler, we could talk about whether the ticks are 1 mm apart – or 1/16″ – or whatever. We talk about the resolution of the ruler in terms of distance between ticks. However, if we are going to be more general, we can talk about the distance between two ticks being one “quantum” – a fancy word for the smallest step size on the ruler.
So, when you’re “rounding off to the nearest value” you are “quantising” the measurement (or “quantizing” it, if you live in Noah Webster’s country and therefore you harbor the belief that wordz should be spelled like they sound – and therefore the world needz more zees). This also means that the amount of error that you get as a result of that “rounding off” is called “quantisation error“.
In some explanations of this problem, you may read that this error is called “quantisation noise”. However, this isn’t always correct. This is because if something is “noise” then is is random, and therefore impossible to predict. However, that’s not strictly the case for quantisation error. If you know the signal, and you know the quantisation values, then you’ll be able to predict exactly what the error will be. So, although that error might sound like noise, technically speaking, it’s not. This can easily be seen in Figures 1 through 3 which demonstrate that the quantisation error causes a periodic, predictable error (and therefore harmonic distortion), not a random error (and therefore noise).
Sidebar: The reason people call it quantisation noise is that, if the signal is complicated (unlike a sine wave) and high in level relative to the quantisation levels – say a recording of Britney Spears, for example – then the distortion that is generated sounds “random-ish”, which causes people to just to the conclusion that it’s noise.
Now, let’s talk about perception for a while… We humans are really good at detecting patterns – signals – in an otherwise noisy world. This is just as true with hearing as it is with vision. So, if you have a sound that exists in a truly random background noise, then you can focus on listening to the sound and ignore the noise. For example, if you (like me) are old enough to have used cassette tapes, then you can remember listening to songs with a high background noise (the “tape hiss”) – but it wasn’t too annoying because the hiss was independent of the music, and constant. However, if you, like me, have listened to Bob Marley’s live version of “No Woman No Cry” from the “Legend” album, then you, like me, would miss the the feedback in the PA system at that point in the song when the FoH engineer wasn’t paying enough attention… That noise (the howl of the feedback) is not noise – it’s a signal… Which makes it just as important as the song itself. (I could get into a long boring talk about John Cage at this point, but I’ll try to not get too distracted…)
The problem with the signal in Figure 2 is that the error (shown in Figure 3) is periodic – it’s a signal that demands attention. If the signal that I was sending into the quantisation system (in Figure 1) was a little more complicated than a sine wave – say a sine wave with an amplitude modulation – then the error would be easily “trackable” by anyone who was listening.
So, what we want to do is to quantise the signal (because we’re assuming that we can’t make a better “ruler”) but to make the error random – so it is changed from distortion to noise. We do this by adding noise to the signal before we quantise it. The result of this is that the error will be randomised, and will become independent of the original signal… So, instead of a modulating signal with modulated distortion, we get a modulated signal with constant noise – which is easier for us to ignore. (It has the added benefit of spreading the frequency content of the error over a wide frequency band, rather than being stuck on the harmonics of the original signal… but let’s not talk about that…)
Let’s take a look at an example of this from an equivalent world – digital photography.
The photo in Figure 4 is a black and white photo – which actually means that it’s comprised of shades of gray ranging from black all the way to white. The photo has 272,640 individual pixels (because it’s 640 pixels wide and 426 pixels high). Each of those pixels is some shade of gray, but that shading does not have an infinite resolution. There are “only” 256 possible shades of gray available for each pixel.
So, each pixel has a number that can range from 0 (black) up to 255 (white).
If we were to zoom in to the top left corner of the photo and look at the values of the 64 pixels there (an 8×8 pixel square), you’d see that they are:
86 86 90 88 87 87 90 91
86 88 90 90 89 87 90 91
88 89 91 90 89 89 90 94
88 90 91 93 90 90 93 94
89 93 94 94 91 93 94 96
90 93 94 95 94 91 95 96
93 94 97 95 94 95 96 97
93 94 97 97 96 94 97 97
What if we were to reduce the available resolution so that there were fewer shades of gray between white and black? We can take the photo in Figure 1 and round the value in each pixel to the new value. For example, Figure 5 shows an example of the same photo reduced to only 4 levels of gray.
Now, if we look at those same pixels in the upper left corner, we’d see that their values are
102 102 102 102 102 102 102 102
102 102 102 102 102 102 102 102
102 102 102 102 102 102 102 102
102 102 102 102 102 102 102 102
102 102 102 102 102 102 102 102
102 102 102 102 102 102 102 102
102 102 102 102 102 102 102 102
102 102 102 102 102 102 102 102
They’ve all been quantised to the nearest available level, which is 102. (Our possible values are restricted to 0, 51, 102, 154, 205, and 255).
So, we can see that, by quantising the gray levels from 256 possible values down to only 6, we lose details in the photo. This should not be a surprise… That loss of detail means that, for example, the gentle transition from lighter to darker gray in the sky in the original is “flattened” to a light spot in a darker background, with a jagged edge at the transition between the two. Also, the details of the wall pillars between the windows are lost.
If we take our original photo and add noise to it – so were adding a random value to the value of each pixel in the original photo (I won’t talk about the range of those random values…) it will look like Figure 6. This photo has all 256 possible values of gray – the same as in Figure 1.
If we then quantise Figure 6 using our 6 possible values of gray, we get Figure 7. Notice that, although we do not have more grays than in Figure 5, we can see things like the gradual shading in the sky and some details in the walls between the tall windows.
That noise that we add to the original signal is called dither – because it is forcing the quantiser to be indecisive about which level to quantise to choose.
I should be clear here and say that dither does not eliminate quantisation error. The purpose of dither is to randomise the error, turning the quantisation error into noise instead of distortion. This makes it (among other things) independent of the signal that you’re listening to, so it’s easier for your brain to separate it from the music, and ignore it.
Addendum: Binary basics and SNR
We normally write down our numbers using a “base 10” notation. So, when I write down 9374 – I mean
9 x 1000 + 3 x 100 + 7 x 10 + 4 x 1
9 x 103 + 3 x 102 + 7 x 101 + 4 x 100
We use base 10 notation – a system based on 10 digits (0 through 9) because we have 10 fingers.
If we only had 2 fingers, we would do things differently… We would only have 2 digits (0 and 1) and we would write down numbers like this:
which would be the same as saying
1 x 16 + 1 x 8 + 1 x 4 + 0 x 2 + 1 x 1
1 x 24 + 1 x 23 + 1 x 22 + 0 x 21 + 1 x 20
The details of this are not important – but one small point is. If we’re using a base-10 system and we increase the number by one more digit – say, going from a 3-digit number to a 4-digit number, then we increase the possible number of values we can represent by a factor of 10. (in other words, there are 10 times as many possible values in the number XXXX than in XXX.)
If we’re using a base-2 system and we increase by one extra digit, we increase the number of possible values by a factor of 2. So XXXX has 2 times as many possible values as XXX.
Now, remember that the error that we generate when we quantise is no bigger than 1/2 of a quantisation step, regardless of the number of steps. So, if we double the number of steps (by adding an extra binary digit or bit to the value that we’re storing), then the signal can be twice as “far away” from the quantisation error.
This means that, by adding an extra bit to the stored value, we increase the potential signal-to-error ratio of our LPCM system by a factor of 2 – or 6.02 dB.
So, if we have a 16-bit LPCM signal, then a sine wave at the maximum level that it can be without clipping is about 6 dB/bit * 16 bits – 3 dB = 93 dB louder than the error. The reason we subtract the 3 dB from the value is that the error is +/- 0.5 of a quantisation step (normally called an “LSB” or “Least Significant Bit”).
Note as well that this calculation is just a rule of thumb. It is neither precise nor accurate, since the details of exactly what kind of error we have will have a minor effect on the actual number. However, it will be close enough.
Once upon a time, when I was a young whipper snapper, studying how to be a recording engineer (which is half of being a tonmeister) I had a textbook on sound recording. There were chapters in there on musical instruments, acoustics, microphones, mixing consoles, magnetic tape, and so on.. There was also a section on something called “digital audio” – but it was a portion of the chapter titled “Noise Reduction”.
Fast-forward a couple of years to 1983 and a new technology hit the market called “Compact Disc” (Here’s a fun fact for impressing people at your next dinner party: The “c” at the end of “disc” means it’s an optical medium. If it were magnetic, it would be a “disk”. So: Compact Disc, but Hard Disk.) Back then, the magazine advertisement read “Perfect Sound. Forever.” Then it hit the real world and the complaints started rolling in from people who believed that they knew things about audio. Some of these complaints were valid, and some were less so… Many of the ones that were valid no longer are, but it’s difficult to un-do a first impression.
Nowadays, it is very likely that almost-all-to-all of the music you listen to has been digital at some point in its life. Even if you’re listening to vinyl, it should not surprise you to know that the master version of the recording you’re hearing was probably stored on a hard disk or passed through a digital mixing console – or at least some of the tracks included some kind of digital processing (say, a guitar pedal or a reverb unit, for example). (I know, I know… There are exceptions. However, if you want to send me anti-digital hate mail you may not do it using a digital communication format such as e-mail. Use an analogue pen to write out your words on a piece of paper and send it to me by post. I look forward to receiving your analogue letters.)
Nowadays, a big part of my “day job” is to test (digital) audio systems to find out what’s wrong with them. So, I thought it would be interesting to do a series of postings that describe the typical kinds of errors that I look for (and find) when I’m digging down into the details.
In order to do this, I’m going to start by being a little redundant and describe the basics of how audio is converted from an analogue signal to a digital one – and hopefully address some of the misconceptions that are associated with this conversion process.
A quick introduction to sound
At the simplest level, sound can be described as a small change in air pressure (or barometric pressure) over short periods of time. If you’d like to have a better and more edu-tain-y version of this statement with animations and pretty colours, you could take 10 minutes to watch this video, for example.
That change in pressure can be “captured” by using a microphone, that is (at the simplest level) a device that has a change in air pressure at its input and a change in electrical voltage at its output. Ignoring a lot of details, we could say that if you were to plot a measurement of the air pressure (at the input of the microphone) over time, and you were to compare it to a plot of the measurement of the voltage (at the output of the microphone) over time, you would see the same curve on the two graphs. This means that the change in voltage is analogous to the change in air pressure.
At this point in the conversation, I’ll make a point to say that, in theory, we could “zoom in” on either of those two curves shown in Figure 1 and see more and more details. This is like looking at a map of Canada – it has lots of crinkly, jagged lines. If you zoom in and look at the map of Newfoundland and Labrador, you’ll see that it has finer, crinkly, jagged lines. If you zoom in further, and stand where the water meets the shore in Trepassey and take a photo of your feet, you could copy it to draw a map of the line of where the water comes in around the rocks – and your toes – and you would wind up with even finer, crinkly, jagged lines… You could take this even further and get down to a microscopic or molecular level – but you get the idea… The point is that, in theory, both of the plots in Figure 1 have infinite resolution, both in time and in air pressure or voltage.
Now, let’s say that you wanted to take that microphone’s output and transmit it through a bunch of devices and wires that, in theory, all do nothing to the signal. Let’s say, for example, that you take the mic’s output, send it through a wire to a box that makes the signal twice as loud. Then take the output of that box and send it through a wire to another box that makes it half as loud. You take the output of that box and send it through a wire to a measuring device. What will you see? Unfortunately, none of the wires or boxes in the chain can be perfect, so you’ll probably see the signal plus something else which we’ll call the “error” in the system’s output. We can call it the error because, if we measure the input voltage and the output voltage at any one instant, we’ll probably see that they’re not identical. Since they should be identical, then the system must be making a mistake in transmitting the signal – so it makes errors…
Pedantic Sidebar: Some people will call that error that the system adds to the signal “noise” – but I’m not going to call it that. This is because “noise” is a specific thing – noise is random – so if it’s not random, it’s not noise. Also, although the signal has been distorted (in that the output of the system is not identical to the input) I won’t call it “distortion” either, since distortion is a name that’s given to something that happens to the signal because the signal is there. (We would probably get at least some of the error out of our system even if we didn’t send any audio into it.) So, we could be slightly geeky and adequately vague and call the extra stuff “Distortion plus noise” but not “THD+N” – which stands for “Total Harmonic Distortion Plus Noise” – because not all kinds of distortion will produce a harmonic of the signal… but I’m getting ahead of myself…
So, we want to transmit (or store) the audio signal – but we want to reduce the noise caused by the transmission (or storage) system. One way to do this is to spend more money on your system. Use wires with better shielding, amplifiers with lower noise floors, bigger power supplies so that you don’t come close to their limits, run your magnetic tape twice as fast, and so on and so on. Or, you could convert the analogue signal (remember that it’s analogous to the change in air pressure over time) to one that is represented (and therefore transmitted or stored) digitally instead.
What does this mean?
Conversion from analogue to digital and back
(but skipping important details)
IMPORTANT: If you read this section, then please read the following postings as well. This is because, in order to keep things simple to start, I’m about to leave out some important details that I’ll add afterwards. However, if you don’t add the details, you could (understandably) jump to some incorrect conclusions (that many others before you have concluded…) So, if you don’t have time to read both sections, please don’t read either of them.
In the example above, we made a varying voltage that was analogous to the varying air pressure. If we wanted to store this, we could do it by varying the amount of magnetism on a wire or a coating on a tape, for example. Or we could cut a wiggly groove in a bit of vinyl that has a similar shape to the curve in the plots in Figure 1. Or, we could do something else: we could get a metronome (or a clock) and make a measurement of the voltage every time the metronome clicks, and write down the measurements.
For example, let’s zoom in on the first little bit of the signal in the plots in Figure 1
We’ll then put on a metronome and make a measurement of the voltage every time we hear the metronome click…
We can then keep the measurements (remembering how often we made them…) and write them down like this:
We can store this series of numbers on a computer’s hard disk, for example. We can then come back tomorrow, and convert the measurements to voltages. First we read the measurements, and create the appropriate voltage…
We then make a “staircase” waveform by “holding” those voltages until the next value comes in.
All we need to do then is to use a low-pass filter to smooth out the hard edges of the staircase.
So, in this example, we’ve gone from an analogue signal (the red curve in Figure 3) to a digital signal (the series of numbers), and back to an analogue signal (the red curve in Figure 7).
In some ways, this is a bit like the way a movie works. When you watch a movie, you see a series of still photographs, probably taken at a rate of 24 pictures (or frames) per second. If you play those photos back at the same rate (24 fps or frames per second), you think you see movement. However, this is because your eyes and brain aren’t fast enough to see 24 individual photos per second – so you are fooled into thinking that things on the screen are moving.
However, digital audio is slightly different from film in two ways:
- The sound (equivalent to the movement in the film) is actually happening. It’s not a trick that relies on your ears and brain being too slow.
- If, when you were filming the movie, something were to happen between frames (say, the flash of a gunshot, for example) then it would never be caught on film. This is because the photos are discrete moments in time – and what happens between them is lost. However, if something were to make a very, very short sound between two samples (two measurements) in the digital audio signal – it would not be lost. This is because of something that happens at the beginning of the chain that I haven’t described… yet…
However, there are some “artefacts” (a fancy term for “weird errors”) that are present both in film and in digital audio that we should talk about.
The first is an error that happens when you mess around with the rate at which you take the measurements (called the “sampling rate”) or the photos (called the “frame rate”) – and, more importantly, when you need to worry about this. Let’s say that you make a film at 24 fps. If you play this back at a higher frame rate, then things will move very quickly (like old-fashioned baseball movies…). If you play them back at a lower frame rate, then things move in slow motion. So, for things to look “normal” you have to play the movie at the same rate that it was filmed. However, as longs no one is looking, you can transfer the movie as fast as you like. For example, if you wanted to copy the film, you could set up a movie camera so it was pointing at a movie screen and film the film. As long as the movie on the screen is running in sync with the camera, you can do this at any frame rate you like. But you’ll have to watch the copy at the same frame rate as the original film…
The second is an easy artefact to recognise. If you see a car accelerating from 0 to something fast on film, you’ll see the wheels of the car start to get faster and faster, then, as the car gets faster, the wheels slow down, stop, and then start going backwards… This does not happen in real life (unless you’re in a place lit with flashing lights like fluorescent bulbs or LED’s). I’ll do a posting explaining why this happens – but the thing to remember here is that the speed of the wheel rotation that you see on the film (the one that’s actually captured by the filming…) is not the real rotational speed of the wheel. However, those two rotational speeds are related to each other (and to the frame rate of the film). If you change the real rotational rate or the frame rate, you’ll change the rotational rate in the film. So, we call this effect “aliasing” because it’s a false version (an alias) of the real thing – but it’s always the same alias (assuming you repeat the conditions…) Digital audio can also suffer from aliasing, but in this case, you put in one frequency (which is actually the same as a rotational speed) and you get out another one. This is not the same as harmonic distortion, since the frequency that you get out is due to a relationship between the original frequency and the sampling rate, so the result is almost never a multiple of the input frequency.
Some details that I left out…
One of the things I said above was something like “we measure the voltage and store the results” and the example I gave was a nice series of numbers that only had 4 digits after the decimal point. This statement has some implications that we need to discuss.
Let’s say that I have a thing that I need to measure. For example, Figure 8 shows a piece of metal, and I want to measure its width.
Using my ruler, I can see that this piece of metal is about 57 mm wide. However, if I were geeky (and I am) I would say that this is not precise enough – and therefore it’s not accurate. The problem is that my ruler is only graduated in millimetres. So, if I try to measure anything that is not exactly an integer number of mm long, I’ll either have to guess (and be wrong) or round the measurement to the nearest millimetre (and be wrong).
So, if I wanted you to make a piece of metal the same width as my piece of metal, and I used the ruler in Figure 8, we would probably wind up with metal pieces of two different widths. In order to make this better, we need a better ruler – like the one in Figure 9.
Figure 9 shows a vernier caliper (a fancy type of ruler) being used to measure the same piece of metal. The caliper has a resolution of 0.05 mm instead of the 1 mm available on the ruler in Figure 8. So, we can make a much more accurate measurement of the metal because we have a measuring device with a higher precision.
The conversion of a digital audio signal is the same. As I said above, we measure the voltage of the electrical signal, and transmit (or store) the measurement. The question is: how accurate and precise is your measurement? As we saw above, this is (partly) determined by how many digits are in the number that you use when you “write down” the measurement.
Since the voltage measurements in digital audio are recorded in binary rather than decimal (we use 0 and 1 to write down the number instead of 0 up to 9) then we use Binary digITS – or “bits” instead of decimal digits (which are not called “dits”). The number of bits we have in the number that we write down (partly) determines the precision of the measurement of the voltage – and therefore (possibly), our accuracy…
Just like the example of the ruler in Figure 8, above, we have a limited resolution in our measurement. For example, if we had only 4 bits to work with then the waveform in 4 – the one we have to measure – would be measured with the “ruler” shown on the left side of Figure 10, below.
When we do this, we have to round off the value to the nearest “tick” on our ruler, as shown in Figure 11.
Using this “ruler” which gives a write-down-able “quantity” to the measurement, we get the following values for the red staircase:
When we “play these back” we get the staircase again, shown in Figure 12.
Of course, this means that, by rounding off the values, we have introduced an error in the system (just like the measurement in Figure 8 has a bigger error than the one in Figure 9). We can calculate this error if we just subtract the original signal from the output signal (in other words, Figure 12 minus Figure 10) to get Figure 13.
In order to improve our accuracy of the measurement, we have to increase the precision of the values. We can do this by adding an extra digit (or bit) to the number that we use to record the value.
If we were using decimal numbers (0-9) then adding an extra digit to the number would give us 10 times as many possibilities. (For example, if we were using 4 digits after the decimal in the example at the start of this posting, we have a total of 10,000 possible values – 0.0000 to 0.9999. If we add one more digit, we increase the resolution to 100,000 possible values – 0.00000 to 0.99999 ).
In binary, adding one extra digit gives us twice as many “ticks” on the ruler. So, using 4 bits gives us 16 possible values. Increasing to 5 bits gives us 32 possible values.
If you’re listening to a CD, then the individual measurements of each voltage – the “sample values” – are stored with 16 bits, which means that we have 65,536 possible values to pick from.
Remember that this means that we have more “ticks” on our ruler – but we don’t necessarily increase its range. So, for example, we’re still measuring a voltage from -1 V to 1 V – we just have more and more resolution to do that measurement with.
Finally! We get to the beginning of the point of the posting in the first place. My whole reason for starting this series of postings was to talk about errors in digital audio.
So, the first one to talk about is whether we have “bit matching” in a system where we expect to do so. For example, if you look at the S/P-DIF output of a good-old-fashioned CD player, do the sample values that are transmitted on that wire identical to the ones on the disc?
This is a fairly easy test to make (in theory). All you have to do is to record the digital signal on the S/P-DIF output of your CD player, subtract the original signal that’s on the disc (making sure that you have done your time alignment correctly). If you have anything other than nothing left over, then something went wrong somewhere.
If the result of this test is that you do NOT get nothing remaining, you cannot jump in head first and say that your S/P-DIF output is not working properly. For example, some sound cards have a sampling rate converter at their digital input. So, if you are capturing the CD player’s output using such a sound card on your computer, then perhaps the errors that you see are being produced by your sound card – and not your player.
A little associated story
This was a method that I used to do the final testing of Wireless Power Link for B&O. I created a little software application that made a signal and sent it out digitally to a Wireless Power Link transmitter (which was running with a resolution of 24 bits – giving us 16,777,216 possible values). I then connected a Wireless Power Link receiver’s output to the same computer. The computer knew how much time it took the signal to get from its output, through the wireless transmission system, back to its input (about 5 ms). So, I took the “output” signal, delayed it by that amount, and then subtracted it from the “input” signal. I then made a detector that counted every bit (instead of every sample) that was incorrect.
The reason I was counting bit errors instead of sample errors was that we wanted to be able to diagnose problems if we found them. If you find out that “this sample is wrong” – you don’t necessarily know whether it was one or more bit errors that caused the problem. By counting bit errors, you have a little more information that can help you diagnose the source of problems when you find them.
Sidebar: since this test was running at 48 kHz and 24 bits with a 2-channel system, that means that there were 2,304,000 bits per second being checked every second
This test ran 24-hours a day continuously for over 11 days. In that time, we found 0 bit errors. That means that we got 0 errors in more than 2,189,721,600,000 bits, which was good.
Now, just before anyone gets excited: that test was run to find out whether the WPL system was able to deliver a bit-perfect output in the absence of any external disturbances. So, the transmitter and the receiver were not moved at any time during the test, and nothing was moved between them – and the result was that the system behaved perfectly.
I’m often asked about my opinion regarding sound quality vs. compression formats or sampling rates or bit depths or psychoacoustic CODEC’s or other things like that…
Of course, there are lots of ways to decide on such an opinion, depending on what parameters you use to define “sound quality” and therefore what it is you’re asking specifically…
One way to think of this is to consider that the original sound file is the “reference” (regardless of how “good” or “bad” it is…), and when you encode it somehow (say, by changing sampling rates, or making it an MP3 file, for example), AND that encoding makes it different, then the resulting difference from the original can be considered an error.
So, I took a compilation of tracks that I often use for listening to loudspeakers. This is about 13 minutes long and is made of excerpts of many different recordings and recording styles, ranging from anechoic female speech, through a cappella choral, orchestral music, jazz, hard rock, heavy metal, and hip hop. The original tracks were all taken from 44.1 kHz / 16-bit CD’s, and the compilation is a 44.1 kHz / 16 bit result. This is what we’ll call the “reference”.
I then used LAME to encode the compilation in different bitrates of MP3. I re-encoded as 320, 256, and 128 CBR (Constant Bit Rate). I also used the “–preset” option to make encodings in the “insane”, “extreme”, “standard”, and “medium” settings (I’ve included the details of this at the bottom in the “Appendix”). Three of these four presets are VBR – the “Insane” setting is a CBR 320 kbps with some tweaked parameters.
I decoded those MP3 files back to PCM, and compared them to the original, of course making sure that everything was time- and gain-aligned. (There are some small differences in the overall level of the original file and the MP3 output – which is different for different bitrates. If I did not do this, then I would be exaggerating the differences between the original and the encoded versions – so this gain difference was calculated and compensated for, before subtracting the original from the MP3.)
Let’s take a look at a plot of the sample values in the left channel of the beginning of the track.
The plot above shows the first 44100 samples in the track (the first second of sound). The red plot is the decoded 128 kbps MP3. The black plot (which is difficult to see because it is overlapped by the red plot – except in the signal peaks) is the original file. For example, if I zoom into the area around the beginning of the sound (say, starting around sample number 15800) then we see this
So, as you can see in the two plots above, the decoded 128 kbps MP3 and the original 44.1/16 file are different. But, the difference is small relative to the levels of the signals themselves. The question is, how small is the difference, exactly?
We can find this out by subtracting the original signal from the decoded MP3 output, sample by sample. The result of this is shown in the plot below.
Notice that the vertical scale of the plot in Figure 3 is small. This is because it shows the difference between the two lines in Figure 2, which is also quite small.
Let’s think for a minute about how I arrived at the signal in Figure 3. I subtracted the Original signal from the MP3 output. In other words:
MP3 output – Original = Difference
If we consider that the difference between the MP3 output and the Original can be thought of as an “error”, and if I move the terms in the equation above, I get the following:
MP3 output – Original = Error
Original + Error = MP3 output
So, the question is: how loud is that error relative to the signal we’re listening to? The idea here is that, the louder the error, the easier it will be to detect.
Figure 4, below, shows this level difference over time. The black curve is a running RMS level of the decoded 128 kbps MP3 file. As you can see there, it ranges from about -30 dB FS to about +10 dB FS. You may think that it’s strange that it “only” goes to -10 dB FS – but this is because the time window I’m using to calculate the RMS value of the signal is 500 ms long. The peaks of the track reach full scale, but since my time window is long, this tends to pull down the apparent level (because the peaks are short). (NB: If you want to argue about the choice of a 500 ms time window, please wait until I’ve followed up this posting with another one that divides things up by frequency band…)
The res curve in Figure 4 is a running RMS value of the Error signal – the difference between the MP3 file and the original. As you can see there, that error signal ranges from about -50 dB FS to about -30 dB FS, give or take…
We can find the running value of the difference between the level of the MP3 file and the level of the Error it contains by subtracting the black curve from the red curve. The result of this is shown in Figure 5, below.
So, Figure 5, therefore, shows the measure of how loud the signal is relative to the error that makes it different from the original. If this error signal were just harmonic distortion, then we could call this a measure of THD in dB. If it were just good-old-fashioned noise, like on a magnetic tape, then we could call it a signal-to-noise ratio. However, this is neither distortion or noise in the traditional sense – or, maybe more accurately, it’s both…
So, let’s call the plot in Figure 5 a “signal-to-error ratio”. What we can see there is that, for this particular track, for the settings that I used to make the 128 kbps MP3 file, the error – the MP3 artefacts – are only 20 to 25 dB below the signal most of the time. Now, don’t jump to conclusions here. This does not mean that they would be as audible as white noise that is only 25 dB below the signal. This is because part of the “magic” of the MP3 encoder is that it tries to ensure that the error can “hide” under the signal by placing the error signal in the same frequency band(s) as the signal. Typically, white noise is in a different band than the signal, so it’s easier to hear because it’s not masked. So, be very careful about interpreting this plot. This is a measurable signal-to-error ratio, but it cannot be directly compared to a signal-to-noise ratio.
Let’s now increase the bitrate of the MP3 encoding, allowing the encoder to increase the quality.
Figure 6 and 7 show the same information as before, but for a 256 kpbs encoding of the same track. As you can see there, by doubling the bitrate of the MP3, we have increased our signal-to-error ratio by about 10 to 15 dB or so – to about 35 or 40 dB.
As you can see in Figures 8 and 9 above, increasing the MP3 bitrate to 320 kbps can improve the Signal-to-Error ratio from about 25 dB (for 128 kbps) to about 40 dB or so.
Now, if you’re looking carefully, you might notice that, some times in the track that I used for testing, the signal-to-error ratio is actually worse for the 320 kbps file than it is for the 256 kbps file – all other things being equal in the LAME converter parameters. This is a bit misleading, since what you cannot see there is the frequency spectrum of the error signal. I’ll deal with that in a future posting – with some more analysis and explanation to go with it.
For now, let’s play with the VBR presets in LAME. I’ll just show the signal-to-error plots for the 4 settings.
So, as you can see in Figures 10 through 13, the signal-to-error ratio can be improved with the VBR presets, reaching a peak of over 60 dB for the “Insane” setting, for this track…
As I said a couple of times above:
- You have to be careful about interpreting these graphs from a background of “knowing” what a SNR is… This error is not normal “distortion” or “noise” – at least from a perceptual point of view…
- I’ll go further with this, including some frequency-dependent information in a future posting.
Appendix – LAME parameters and verbose output
For the geeks…