What is a good resource on bit depths?

[Dale-c]

I am curious to learn more about the differnt bit depths. 8bit really itrigues me because I tried bouncing to 8 bit just to see what it would sound like and I found it odd that I can hear the full frequency range (as I assumme is because I still had it at 48k) and also dynamics were still pretty good but it sounded like FM radio with static, Does anyone know what causes this? I don't understand why it would sound like all of the audio info is still there but with added noise. I post some audio to the internet every week and I am always on the lookout to inprove the quality for the file size. I currently use both mp3 at 16kbps and AAC at 32 kbps, both mono. I am wondering if I can get better frequency response by going with a higher sampling rate (such as 32k) but lowering the bit rate to either 8 or 12.

[O.G. Killa]

Dither Explained (White Paper)

Not sure if this is what you're looking for but it might give you a better understanding of how Bit Depth affects sound quality and why there is noise when going from higher bit depths to lower ones.

[JasonRomney]

Resolution (bit depth) has little if nothing to do with frequency response. That is determined by the sampling rate. Without going into too much detail, the basic rule is that the sampling rate has to be at least twice that of the highest frequency that is being reproduced in order to even come close. Since 20 kHz is generally accepted as the highest frequency that most people can hear, the sampling rate of 44.1 kHz found in CD audio is adequate. Really more so since there isn't really much going on between 16 and 20 kHz anyway. Pretty much all good A/D converters band limit the audio to 20 kHz before it gets digitized. Anything higher than 44.1 kHz is just making the digital approximation of the frequency even more accurate.

Resolution determines the dynamic range of the digital approximation. The higher the resolution (bit depth) the wider the dynamic range and the more accurate the digital system is at interpreting and recreating the differences in level of your recordings. The dynamic range of 8 bit audio is around 45 dB which if used effectively is adequate when you consider that the with an average noise floor of 60 dB SPL of most listening environments, which leaves you with around 45 dB to work with before you start approaching the threshold of pain. 16 bit resolution gives you over 90 dB of dynamic range but very few recording engineers really use all of that. The problem goes back to that whole signal to noise ratio of most listening environments. Most of your recording work is done with levels in the upper 40 dB of your 16 bit range which suddenly throws out most of the bits in that 90+ dB. So although no one ever uses the entire dynamic range of 24 bit digital audio, it does provide you with a much higher resolution within the 40 dB or so that you're actually using.

Most of the artifacts you hear when converting your 16 bit stuff to 8 bit is caused by levels that fall in place on the 16 bit scale but have no equivalent value in the 8 bit scale so they get bumped to whatever is closest. This can often cause some nasty artifacts. Dither can help with that if it is used properly.

That's all I have time to write now but let me know if you have any more questions.

[Dale-c]

I am familiar with the sampling rates and needing to be twice the highest frequency etc. The added info was very interesting about the bits though. I was somewhat familiar with that but now it is a little clearer. I use dither to go from 24 bit down to 16 when I bouce a session but do you know of a dither that will go from 16 down to 8? Also, is using 8 bit for the web even practicle? I am just looking for a way to get higher sample rates without increasing my bit rate of 16kbps and 32 kbps which is of course pretty low quality.
Thanks for the info

[Dale-c]

One more thing, does anyone know of any software that will dither a 16 bit file to an 8 bit file? Digi's seems to only go from 24 t o16

[JasonRomney]

The Waves Ultramaximizer will do an 8 bit dither.

Be careful not to get the bit rate of an mp3 confused with the bit depth of a standard PCM file. They are two very different things.

Bit rate is the amount of data that is being moved every second. This term is much more common when you're talking about internet distribution. How many bits can we send over the wire at a time?

The bit rate of PCM audio 16 bit, 44.1 kHz is 705.6 kb per second. I'm not familiar enough with mp3 encoding to go into much detail but the crux of it is that it looks at your file and attempts to prioritize the information. It looks at each bit and ranks it according to how important it is in reproducing the right sound. It then starts from the bottom, eliminating the least important information until it eliminates enough bits to fit within the limit (128 kb/s, 64 kb/s, etc). There's more to it than that but that's the basic idea as I understand it. So if you're really looking to crunch things, it stands to reason that the smaller you can get your file before you encode it (either by dropping the resolution or sampling rate) the less bits will need to be eliminated by the encoder and you can get away with much lower bit rates. I've never been concerned enough about the size of the files I post on the internet to try this theory but it seems logical.

Anyone disagree or have any further insight?

[Duardo]

Without going into too much detail, the basic rule is that the sampling rate has to be at least twice that of the highest frequency that is being reproduced in order to even come close.

That's not quite right...it's either reproduced correctly or it's not. Sure, no converter is perfect...but that's because the audio has to pass through a bunch of analog circuitry before the conversion, then get processed further once converted, and the same thing has to happen in reverse upon conversion from digital back to analog. The filters, etc that are used in these processes aren't perfect, but they're getting better all the time. But all that you need to perfectly recreate a frequency is to sample it at a rate that's more than twice that frequency.

Anything higher than 44.1 kHz is just making the digital approximation of the frequency even more accurate.

That's not quite right either. All it does is allow you to capture frequencies that are higher than 20kHz. To be fair, it does also allow more room for imperfections in terms of the filters used to get rid of the information we don't want, but the conversion process itself doesn't get any more accurate below 20kHz when the sampling rate goes up above 44.1kHz.

So although no one ever uses the entire dynamic range of 24 bit digital audio, it does provide you with a much higher resolution within the 40 dB or so that you're actually using.

Actually, it doesn't. If you're only using 40dB of dynammic range, 16 bits is way more than sufficient. The only extra "resolution" you're gaining is in well below the level of your lowest signal. The nice thing about 24-bit recording (if you're only using a small portion of its dynamic range) is that you don't need to worry about setting your levels so that you're within a dB or two of 0 dBFS. With material with a 40-dB dynamic range, in a 24-bit system you can easily peak at -15dB or -20dB with no loss of quality at all.

-Duardo

[JasonRomney]

That's not quite right...it's either reproduced correctly or it's not.

Actually the accuracy varies. If you are trying to digitize a sine wave with a sampling rate that is only twice the frequency, you get the same frequency but it will be a square wave instead of a sine wave since you only have two samples. Now the D/A converter will try to smooth that out but it will be an approximation at best and if you do any processing on it in the digital realm, you'll be processing a square wave rather than a sine wave. More samples are required to start approximating the curves.

but the conversion process itself doesn't get any more accurate below 20kHz when the sampling rate goes up above 44.1kHz

Also not true. It comes back to the curves thing. A digitized sound wave is not smooth like the original analog wave. It's a series of steps representing the level captured for each sample. The more samples per second, the smaller the steps, the smoother the wave, and the less work the smoothing filter on the DAC has to do to smooth the wave out. The more detail you give it to work with, the closer it is going to get to the smooth analog equivalent. Now, yes, higher sampling rates allow you to adjust the roll-off of the smoothing filters so they're outside of the audio range but the frequencies within the audible range certainly benefit as well from the increased detail.

This is much easier to understand with illustrations. Here's a good website that has some pretty good explanations and images.

http://www.indiana.edu/~emusic/digital_audio.html

I agree with you on the resolution thing. I think we're just using different words to describe the same thing. As you say, with 24 bit audio, you don't need to be as concerned with your levels because you can get an accurate recording without trying to use the entire range. This is due to the added resolution along the scale. You can get the same if not better resolution using only a portion of the 24 bit scale than you get trying to use a larger portion of the 16 bit scale by setting your levels as high as you can before clipping.

[Duardo]

Actually the accuracy varies. If you are trying to digitize a sine wave with a sampling rate that is only twice the frequency, you get the same frequency but it will be a square wave instead of a sine wave since you only have two samples. Now the D/A converter will try to smooth that out but it will be an approximation at best and if you do any processing on it in the digital realm, you'll be processing a square wave rather than a sine wave. More samples are required to start approximating the curves.

That's a common misconception, and it certainly makes sense on an intuitive level, but it's wrong. What you sample may appear to be a square wave when you look at it on a screen, but what comes out of your D/A converter is a sine wave. The D/A converter doesn't "connect the dots"; what it does is fit a curve to the samples it's given, and there is one and only one curve that can be fit to those "dots" since everything above the Nyquist frequency has been filtered out. It does not get any more accurate when you raise the sampling rate. If it did, then the Nyquist theorem would be incorrect and digital audio as we know it would not work. The problems we may hear below 20kHz are related to imperfections in converter design, not the sampling rate itself.

There are some digital processes that do work better at higher sampling rates, but 44.1kHz audio can easily be upsampled for processing and then downsampled. There are hardware pieces and software plugins that have been working that way for years.

A digitized sound wave is not smooth like the original analog wave. It's a series of steps representing the level captured for each sample. The more samples per second, the smaller the steps, the smoother the wave, and the less work the smoothing filter on the DAC has to do to smooth the wave out.

True, a digitized wave isn't smooth like the original wave, but we're not talking about digitized waves. We're talking about the waves that come out of the D/A converter, and once the digital signal has been filtered the resultant wave is and analog wave that's just as smooth as the original wave. As long as it's below the Nyquist frequency, it's just as smooth whether it's represented by two, twenty, two hundred, or two thousand samples.

The more detail you give it to work with, the closer it is going to get to the smooth analog equivalent. Now, yes, higher sampling rates allow you to adjust the roll-off of the smoothing filters so they're outside of the audio range but the frequencies within the audible range certainly benefit as well from the increased detail.

With today's converters, which filter very gently in the analog domain, sample at an extremely high frequency, and then filter much more agressively (but much more transparently) in the digital domain, that's simply not the issue it was a few short years ago.

The University of Indiana link is great, but doesn't say anywhere that there's any advantage to sampling at a rate that's higher than twice the frequency you want to capture.

-Duardo

[JasonRomney]

The D/A converter doesn't "connect the dots"; what it does is fit a curve to the samples it's given, and there is one and only one curve that can be fit to those "dots" since everything above the Nyquist frequency has been filtered out.

Okay but what about different types of waves. I mean, normally we're dealing with much more complex waves but for purposes of discussion we can limit it to simple tones. There are many different ways for the DAC to match a curve to the set of samples. If you've only got two samples for a cycle, that's one at the high peak and one at the low peak. So that sound could be a sine wave, a triangle wave, a sawtooth wave, or a square wave. How does the DAC know which wave it should be? The only information it has is the two sample points. The digital data is going to look the same to the DAC whether it is supposed to be a sine or a triangle. If you've got more samples, it's going to have a better idea what type of smoothing to apply.

Here's an image from the Indiana link. It shows the source complex wave of 3 kHz (with a lot of harmonics) being digitized at 9 kHz and at 48 kHz. You can't tell me that the top one is going to come out of the DAC as accurate as the bottom one. There's just not enough information there. How is the DAC supposed to assume all those sublte changes along the wave if it's dealing with only three samples? So yeah, the higher sampling rate of 48 kHz is allowing the harmonics to be interpreted but what about the harmonics at the very top of the frequency spectrum? They're the ones that are still suffering even at 48 kHz (which is much more than twice the highest frequency). A sampling rate higher than the 48 kHz will allow those harmonics (that are still within the audio range) to be interpreted more correctly.

An interesting experiment would be to put an oscilloscope on each end of a digital system and feed different types of waves into it using a sampling rate at twice the frequency and see how close each type of wave (sine, triangle, sawtooth, etc) at the end gets to the original wave. Anybody done this before?