Increasing bit depth in a fixed-point system
This is the first of a series of postings about strategies for converting from one bit depth to another, including conversion back and forth between fixed point and floating point encoding. It’ll be focusing on a purely practical perspective, with examples of why you need to worry about these things when you’re doing something like testing audio devices or transmission systems.
As we go through this, it might be necessary to do a little review, which means going back and reading some other postings I’ve done in the past if some of the concepts are new. I’ll link back to these as we need them, rather than hitting you with them all at once.
To start, if you’re not familiar with the concept of quantisation and bit depth in an LPCM audio signal, I suggest that you read this posting.
Now that you’re back, you know that if you’re just converting a continuous audio signal to a quantised LPCM version of it, the number of bits in the encoded sample values can be thought of as a measure of the system’s resolution. The more bits you have, the more quantisation steps, and therefore the better the signal to noise ratio.
However, this assumes that you’re using as many of the quantisation steps as possible – in other words, it assumes that you have aligned levels so that the highest point in the audio signal hits the highest possible quantisation step. If your audio signal is 6 dB lower than this, then you’re only using half of your available quantisation values. In other words, if you have a 16-bit ADC, and your audio signal has a maximum peak of -6 dB FS Peak, then you’ve done a 15-bit recording.
But let’s say that you already have an LPCM signal, and you want to convert it to a larger bit depth. A very normal real-world example of this is that you have a 16-bit signal that you’ve ripped from a CD, and you export it as a 24-bit wave file. Where do those 8 extra bits come from and where do they go?
Generally speaking, you have 3 options when you do this “conversion”, and the first option I’ll describe below is, by far the most common method.
Option 1: Zero padding
Let’s simplify the bit depths down to human-readable sizes. We’ll convert a 3-bit LPCM audio signal (therefore it has 2^3 = 8 quantisation steps) into a 5-bit representation (2^5 = 32 quantisation steps), instead of 16-bit to 24-bit. That way, I don’t have to type as many numbers into my drawings. The basic concepts are identical, I’ll just need fewer digits in this version.
The simplest method to do is to throw some extra zeros on the right side of our original values, and save them in the new format. A graphic version of this is shown in Figure 1.
There are a number of reasons why this is a smart method to use (which also explains why this is the most common method).
The first is that there is no change in signal level. If you have a 0 dB FS Peak signal in the 3-bit world, then we assume that it hits the most-negative value of 100. If you zero-pad this, then the value becomes 10000, which is also the most-negative value in the 5-bit world. If you’re testing with symmetrical signals (like a sinusoidal tone) then you never hit the most-negative value, since this would mean that it would clip on the positive side. This might result in a test that’s worth talking about, since sinusoidal tone that hits 011 and is then converted to 01100. In the 5-bit world, you could make a tone that is a little higher in level (by 3 quantisation levels – those top three dotted lines on the right side of Figure 1), but that difference is very small in real life, so we ignore it. The biggest reason for ignoring this is that this extra “headroom” that you gain is actually fictitious – it’s an artefact of the fact that you typically test signal levels like this with sine tones, which are symmetrical.
The second reason is that this method gives you extra resolution to attenuate the signal. For example, if you wanted to make a volume knob that only attenuated the audio signal, then this conversion method is a good way to do it. (For example, you send a 16-bit digital signal into the input of a loudspeaker with a volume controller. You zero-pad the signal to 24-bit and you now have the ability to reduce the signal level by 141 dB instead of just 93 dB (assuming that you’re using dither…). This is good if the analogue dynamic range of the rest of the system “downstream” is more than 93 dB.) The extra resolution you get is equivalent to 6 dB * each extra bit. So, in the system above:
(5 bits – 3 bits) = 2 extra bits
2 extra bits * 6 dB = 12 dB extra resolution
There is one thing to remember when doing it this way, that you may consider to be a disadvantage. This is the fact that you can’t increase the gain without clipping. So, let’s say that you’re building a digital equaliser or a mixer in a fixed-point system, then you can’t just zero-pad the incoming signal and think that you can boost signals or add them. If you do this, you’ll clip. So, you would have to zero-pad at the input, then attenuate the signal to “buy” yourself enough headroom to increase it again with the EQ or by mixing.
Option 2
The second option is, in essence, the same as the trick I just explained in the previous paragraph. With this method, you don’t ONLY pad the right side of the values with zeros, you pad the values on the left as well with either a 1 or a 0, depending on whether the signals are positive or negative. This means that your “old” value is inserted into the middle of the new value, as shown below in Figure 2. (In this 3- to 5-bit example, this is identical to using option 1, and then dropping the signal level by 6 dB (1 of the 2 bits)).
If your conversion to the bigger bit depth is done inside a system where you know what you’ve done, and if you need room to scale the level of the signal up and down, this is a clever and simple way to do things. There are some systems that did this in the past, but since it’s a process that’s done internally, and we normal people sit outside the system, there’s no real way for us to know that they did it this way.
(For example, I once heard through the grapevine that there was a DAW that imported 24-bits into a 48-bit fixed point processing system, where they padded the incoming files with 12 bits on either side to give room to drop levels on the faders and also be able to mix a lot of full-scale signals without clipping the output.)
Option 3
I only add the third option as a point of completion. This is an unusual way to do a conversion, and I only personally know of one instance where it’s been used. This only means that it’s not a common way to do things – not that NO ONE does it.
In this method, all the padding is done on the left side of the binary values, as shown below in Figure 3.
If we’re thinking along the same lines as in Options 1 and 2, then you could say that this system does not add resolution to attenuate signals, but it does give you the option to make them louder.
However, as we’ll see in Part 2 of this series, there is another advantage to doing it this way…
Nota Bene
I’ve written everything above this line, intentionally avoiding a couple of common terms, but we’ll need those terms in our vocabulary before moving on to Part 2.
If you look at the figures above, and start at the red 0 line, as you go upwards, the increase in signal can be seen as an increase in the left-most bits in each quantisation value. Reading from left-to-right, the first bit tells us whether the value is positive (0) or negative (1), but after this (for the positive values) the more 1s you have on the left, the higher the level. This is why we call them the Most Significant Bits or MSBs. Of course, this means that the last bit on the right is the Least Significant Bit or LSB.
This means that I could have explained Option 1 above by saying:
The three bits of the original signal become the three MSBs of the new signal.
… which also tells us that the signal level will not drop when converted to the 5-bit system.
Or I could have explained Option 3 by saying:
The three bits of the original signal become the three LSBs of the new signal.
.. which also tells us that the signal level will drop when converted to the 5-bit system.
Being able to think in terms of LSBs and MSBs will be helpful later.
Finally… yes, we will talk about Floating Point representations. Later. But if you can’t wait, read this in the meantime.