Down sampling is down sampling, not lossy compression.
If it was the same as lossy compression, then that would imply would data on the CD would be uncompressed on playback to provide some resemblance to the original high sample rate master.
This does not happen on CD, as the missing information from the original master is irretrievably lost. There is no decompression on playback, and so no extra information is generated.
If you take a picture and remove half the pixels, you have not compressed it, you have removed half the pixels. This is equivalent to downsampling. There is no way of getting those missing pixels back.
If you use a compression scheme that allows assign more data to those pixels are more important to the way humans perceive images, you have used lossy compression. You can increase the perceived quality of the image after decompression.
Lossy compression also implies a trade off between human perception and available bandwidth. As perception does not factor in linear PCM audio, you cannot say CD uses lossy compression.
This is inadvisable as it can cause elevator bones and whip flash tones. CD may not be so proficient at reproducing jig-saw jazz and the get-fresh flow, but the lower noise floor and increased resolution allows better jives and jamboree handouts.
Unless you can be sure that the vinyl and CD masters were identical, your tests are invalid. Due to the limitations of vinyl, and the current trend in CD mastering, they are unlikely to be the same masters.
A better comparison is to try comparing the direct output of the turntable to the same through 16bit 44.1KHz ad/da conversion. In double blind tests, no one has ever been able to reliably tell the difference.
What's also interesting is that even for many records that were recorded and mastered entirely analogue, they still have been put through 16/44.1 ad/da conversion. This conversion is necessary because a one revolution digital delay line is used so the variable groove width spacing can be calculated in advance while the record is being cut. It has been common to use digital delays for this since the first decent lexicon ones appeared some time in the early '80s. Before that, they would use a tape delay. Yikes!
Arrgggh. Dither *entirely* prevents quantisation noise. It's not just a case of adding noise as an antidote, it's a fundamental principle of digital recording, an absolute necessity.
One day, try making some 8bit recordings, by dithering from 16bit. You will find that they are noisy, but there is *no* distortion or grittiness! It's not like vinyl, where harmonic distortion increases as the level gets lower.
It's also possible to noise shape the dithering too, so signals can be preserved below the noise floor by shifting the dither noise to less audible parts of the spectrum. That's the icing on the cake though, simple 1/2 lsb is technically correct.
Actually, normalisation is the least extreme form of loudness maximisation. It does not audibly change the signal at all, apart from a volume increase. There is no dynamic range compression of any kind going on. Normalisation to 0dbfs is identical to turning up your master fader until the highest sample peak in the song hits zero. This does not add distortion of any kind, as you are not going over zero. You can normalise to other levels too, it does not have to be zero db.
This is always done as an off-line process, as the audio has to be analysed first to find it's highest peak.
If you want extreme loudness maximisation, you want a fast look ahead peak limiter and multi band compressor.
The RIAA eq transfer curve in the phono preamp will be totally incorrect if you transfer at a different speed. The record will sound strangely eq'd. That's not even considering the damage pitch shifting will do as well.
I suppose that if one is buying that turntable from thinkgeek, then one probably does not care what it sounds like anyway.
The difference is that if a verified watermark is *required* to access the information, then the software to find and compare that watermark, and the customers identity key, has to be on their computer. Therefore there is a way to analyse and reverse engineer that software. It's obvious if you have not cracked it as you cannot access the information.
With music watermarking for tracking users, a valid watermark is not required to access, and the watermark analysis software and private keys are secret and only used by the investigators. The customer cannot check they have broken the watermark as they have no way of verifying whether their attack has been successful.
All the proposed schemes will survive this easily. Testing for resistance to transcoding is the first thing they try, as lossy compression really does screw up audio.
The tests I've seen survived multiple generations too. A bit error rate of 0.022 percent corruption of the retrieved watermark at 64Kbps is typical. Using simple error correction as well makes it rather reliable, and the watermark is embedded by repeating it many times, so you have ample opportunity to recover it.
Anyway, the hardest part, that no one seems to mention, is telling when you have successfully removed the watermark. You can screw with the audio all day, but without the analysis tools you cannot be sure you have removed it.
"If you're choosing low-latency professional audio recording, which one do you pick?"
Jack. As none of the other servers are intended for low latency professional audio recording.
"SAE is behind Ardour? Great. But who's behind them, who's doing the back-end that makes low-latency multitrack possible?"
No one needs to do the back end because it's already been done. I've been doing low latency audio on Linux for about five years or so. The first RT patches for Linux appeared some time in the 2.4 series.
If you want to do professional audio on Linux, you use Jackd. It's as simple as that. There is not a combination of solutions, there is only one solution. If you had used Linux for audio, you would know this already.
One of the reasons for this is that the different sound servers fill different needs. Jackd is callback based and clients run synchronously. This is important for latency, but demands that all the audio apps should be real time safe. The other sound servers are for much less critical situations and work somewhat differently.
On Windows, it's a bit more complicated as there are a number of competing sound standards (ASIO,MME,DirectX,WDM,GSIF,EASI,KStream etc). There is unfortunately no equivalent to Jackd (the Windows port is not finished yet), but you can sort of do some of the same stuff with Rewire.
The constant re-allocation of the available bandwidth among the sub bands, and the way the quantization noise follows it, is part of the problem. It sounds like subtle amplitude modulation of the signal to me, like with a badly set up multiband compressor, and you can hear parts of the ambient noise floor vanishing and re-appearing. Other distortions are caused by the ringing of the FFT filters.
Another sound containing short high energy impulses that MP3 encoders really don't like are clavs. http://ff123.net/preecho.html
I bet I could make a sound combining simultaneous swept sines, clav like impulses and jangling keys that would obviously fox any mp3 encoder. It's not really fair as it's not really music and is designed to reveal artifacts, but it would be a good educational tool.
"An MP3 is simply the same signal that you find on a CD transformed into the frequency domain, frequencies with lesser engery quantized greater, or dropped if below the absolute threshold of hearing, some spatial information discarded (depending on the encoding mode), and written out as a bitstream."
Even though it is wrong, as it ignores temporal masking, one of the key ways in which 'inaudible' information is discarded with MP3 compression. And that the *relationship* between the discarded frequency bands is vital, it's not just a case of doing an FFT and removing any bins with lesser energy. I guess this is not the place for a technical discussion of lossy codecs, but I live in hope.:)
Could you elucidate a little on why you think I have no idea about how lossy audio compression works?
Just stating that and talking about PASC does not really tell me anything. I'm aware it works very differently to a multiband compressor. Some of the artifacts sound quite similar to a badly set one though, and it's the only way I think to communicate what they sound like to me.
The jangling keys test tends to show up problems in audio gear as it contains large amounts of ultrasonic information. Even some standard software equalizers can sound quite bad with it.
The artifacts of MP3 compression, both in the time and frequency domain, are more obvious with uncompressed source material.
Stick a decent mic up in a room, and record some jangling car keys while slowly walking away from the mic till it's almost inaudible, and MP3 will sound awful. (And so will quite a lot of prosumer recording gear recording uncompressed.:)
To me, MP3 artifacts sound like the distortion you get by putting audio through a multiband compressor with a low ratio and threshold, but with too fast attack and release times. So, you get a constant subtle distortion. I doubt most people notice, but as I am used to using multibands, I am used to making sure they are working right, and it sounds odd.
Perhaps with heavily audio compressed material, the artifacts of the MP3 compression are less obvious as it's so smashed to start with.
That's a bit like saying that bad food can be improved by removing your sense of taste though.:)
Thanks for the reply. I thought it would be a cover or pit band or something. There is nothing wrong with playing covers (it's all about making people enjoy themselves after all). You don't really need to have recordings or an established fanbase either, as people know pretty much what they are going to get in advance.
It's not the part of the music industry I'm in. I only work with original acts, for better or for worse. I guess I've been doing it so long I tend to see everything from that perspective.
"Before the 20th century musicians made their money by performing. During the 20th century many musicians made their money by recording music. After the advent of the internet musicians will once again make their money by performing and use their recordings as advertising (as everybody but the RIAA bands do now)."
Piffle. Recorded music is an artform that first appeared in the 20th century, and is distinctly different from acoustic live music. To say people should not be paid for it is to say that only certain musical genres should make money.
Anyway, no one makes any money gigging. You have absolutely no idea how expensive and how time consuming it is. The gigs are to advertise the recordings.
Just work out how much a band with four members and a single roadie/manager/sound guy would have to make every day by gigging to feed everyone and pay the rent. You'd need to be doing 4-5 gigs a week, every week, all year. How long can a small band do that before they exhaust their fanbase? Probably a couple of months. That's assuming their fanbase is not scattered thinly all over the world, and can actually make it to the local venues the band can reach.
Only the larger or RIAA acts can play places large enough for the economies of scale to kick in and start making profit.
"Even a reference tone of 1KHz gives you only 44 and a bit samples per cycle. That is totally inadequate to describe a complex sound in all its detail, whatever Nyquist says."
Ah, but it is adequate. Any 'complexity' not recorded contains frequencies above half the sample rate. Any sine wave can be described as long as you have more than two samples.
"Anyone, even your grandfather, can hear 5.5KHz. 8 samples per cycle. Better than 4 but hardly accurate. Why don't you try reducing a graph of, say, a stock price over even a day to 8 samples, and then translate it back, without losing information."
If the stock price is low pass filtered in the same way as digital audio before conversion, this is not only possible, but trivial to do correctly.
1 bit SACD works with exactly the same theory as multibit PCM audio. They are no different in principal.
No. High sample rates do not remove audible stair stepping in the reconstructed waveform.
There is never stair stepping in the reconstructed waveform, regardless of the sampling frequency used, provided both the input and output are correctly band limited to below 1/2 Fs.
It does not matter if you sample at 44.1Khz, 96Khz, 2MHz of 500Khz.
That is the absolute first principle of digital audio recording.
Remember, samples are not signals. What comes out of the reconstruction filter of a D/A converter is not the same as what you see on your computer monitor waveform display.
There is quantization noise, but this is related to bit depth, and is *completely* inaudible due to the use of 1/2 LSB dither. A greater bit depth gives you a lower noise floor, it does not provide "more room for subtleties to be revealed in louder passages".
I don't expect you to understand this. But don't worry, reading the comments on here it appears that very few people do.
Arrrgh. The slashcode has removed all the 'less than' signs from my post. The first and last times I write '1/2 Fs' should have a 'less than' symbol before them.
Nyquist states that there is *exact* reproduction of all frequencies 1/2 Fs. You only need two and a bit samples to reconstruct a sine and it's phase, and all waveforms can be thought of as a sum of sinewaves of different frequencies and phases. So, as any sine below 1/2 Fs is reconstructed exactly, so is any possible waveform with harmonics 1/2 Fs.
16/44.1 digital audio is already at the threshold of our perception. Ie, it's had to hear the difference between it and 24/96.
Lossy compression uses a perceptual model based on the way we perceive sound.
So, to create a lossy codec that is of a higher quality than 16/44.1, we need to know how we hear the difference between 16/44.1 and higher sample rate/bit depths. The available data bandwidth can then be concentrated on preserving what is most important to the human ear about that difference.
Since very few people can tell the difference, building a perceptual model of it is very hard.
Slashdot Mirror
Down sampling is down sampling, not lossy compression.
If it was the same as lossy compression, then that would imply would data on the CD would be uncompressed on playback to provide some resemblance to the original high sample rate master.
This does not happen on CD, as the missing information from the original master is irretrievably lost. There is no decompression on playback, and so no extra information is generated.
If you take a picture and remove half the pixels, you have not compressed it, you have removed half the pixels. This is equivalent to downsampling. There is no way of getting those missing pixels back.
If you use a compression scheme that allows assign more data to those pixels are more important to the way humans perceive images, you have used lossy compression. You can increase the perceived quality of the image after decompression.
Lossy compression also implies a trade off between human perception and available bandwidth. As perception does not factor in linear PCM audio, you cannot say CD uses lossy compression.
Do you understand the difference between noise and non linearity in the context of digital signal processing?
I'd still like the statue anyway.
This is inadvisable as it can cause elevator bones and whip flash tones.
CD may not be so proficient at reproducing jig-saw jazz and the get-fresh flow, but the lower noise floor and increased resolution allows better jives and jamboree handouts.
Unless you can be sure that the vinyl and CD masters were identical, your tests are invalid.
Due to the limitations of vinyl, and the current trend in CD mastering, they are unlikely to be the same masters.
A better comparison is to try comparing the direct output of the turntable to the same through 16bit 44.1KHz ad/da conversion.
In double blind tests, no one has ever been able to reliably tell the difference.
What's also interesting is that even for many records that were recorded and mastered entirely analogue, they still have been put through 16/44.1 ad/da conversion.
This conversion is necessary because a one revolution digital delay line is used so the variable groove width spacing can be calculated in advance while the record is being cut.
It has been common to use digital delays for this since the first decent lexicon ones appeared some time in the early '80s. Before that, they would use a tape delay. Yikes!
Arrgggh.
Dither *entirely* prevents quantisation noise. It's not just a case of adding noise as an antidote, it's a fundamental principle of digital recording, an absolute necessity.
One day, try making some 8bit recordings, by dithering from 16bit.
You will find that they are noisy, but there is *no* distortion or grittiness!
It's not like vinyl, where harmonic distortion increases as the level gets lower.
It's also possible to noise shape the dithering too, so signals can be preserved below the noise floor by shifting the dither noise to less audible parts of the spectrum. That's the icing on the cake though, simple 1/2 lsb is technically correct.
Actually, normalisation is the least extreme form of loudness maximisation.
It does not audibly change the signal at all, apart from a volume increase.
There is no dynamic range compression of any kind going on.
Normalisation to 0dbfs is identical to turning up your master fader until the highest sample peak in the song hits zero. This does not add distortion of any kind, as you are not going over zero.
You can normalise to other levels too, it does not have to be zero db.
This is always done as an off-line process, as the audio has to be analysed first to find it's highest peak.
If you want extreme loudness maximisation, you want a fast look ahead peak limiter and multi band compressor.
nononononono!
The RIAA eq transfer curve in the phono preamp will be totally incorrect if you transfer at a different speed.
The record will sound strangely eq'd. That's not even considering the damage pitch shifting will do as well.
I suppose that if one is buying that turntable from thinkgeek, then one probably does not care what it sounds like anyway.
The difference is that if a verified watermark is *required* to access the information, then the software to find and compare that watermark, and the customers identity key, has to be on their computer. Therefore there is a way to analyse and reverse engineer that software. It's obvious if you have not cracked it as you cannot access the information.
With music watermarking for tracking users, a valid watermark is not required to access, and the watermark analysis software and private keys are secret and only used by the investigators. The customer cannot check they have broken the watermark as they have no way of verifying whether their attack has been successful.
All the proposed schemes will survive this easily.
Testing for resistance to transcoding is the first thing they try, as lossy compression really does screw up audio.
The tests I've seen survived multiple generations too. A bit error rate of 0.022 percent corruption of the retrieved watermark at 64Kbps is typical. Using simple error correction as well makes it rather reliable, and the watermark is embedded by repeating it many times, so you have ample opportunity to recover it.
Anyway, the hardest part, that no one seems to mention, is telling when you have successfully removed the watermark.
You can screw with the audio all day, but without the analysis tools you cannot be sure you have removed it.
There is no problem with millions of water marks.
Embedding just 32 bits of data gives you enough individual codes for 4.3 billion people.
It's quite easy to hide 32bits in three minutes of audio, even with with massive redundancy to prevent circumvention.
"If you're choosing low-latency professional audio recording, which one do you pick?"
Jack. As none of the other servers are intended for low latency professional audio recording.
"SAE is behind Ardour? Great. But who's behind them, who's doing the back-end that makes low-latency multitrack possible?"
No one needs to do the back end because it's already been done.
I've been doing low latency audio on Linux for about five years or so. The first RT patches for Linux appeared some time in the 2.4 series.
If you want to do professional audio on Linux, you use Jackd. It's as simple as that. There is not a combination of solutions, there is only one solution. If you had used Linux for audio, you would know this already.
One of the reasons for this is that the different sound servers fill different needs. Jackd is callback based and clients run synchronously. This is important for latency, but demands that all the audio apps should be real time safe. The other sound servers are for much less critical situations and work somewhat differently.
On Windows, it's a bit more complicated as there are a number of competing sound standards (ASIO,MME,DirectX,WDM,GSIF,EASI,KStream etc). There is unfortunately no equivalent to Jackd (the Windows port is not finished yet), but you can sort of do some of the same stuff with Rewire.
The constant re-allocation of the available bandwidth among the sub bands, and the way the quantization noise follows it, is part of the problem. It sounds like subtle amplitude modulation of the signal to me, like with a badly set up multiband compressor, and you can hear parts of the ambient noise floor vanishing and re-appearing. Other distortions are caused by the ringing of the FFT filters.
Another sound containing short high energy impulses that MP3 encoders really don't like are clavs.
http://ff123.net/preecho.html
I bet I could make a sound combining simultaneous swept sines, clav like impulses and jangling keys that would obviously fox any mp3 encoder.
It's not really fair as it's not really music and is designed to reveal artifacts, but it would be a good educational tool.
Well, this got +5.
:)
"An MP3 is simply the same signal that you find on a CD transformed into the frequency domain, frequencies with lesser engery quantized greater, or dropped if below the absolute threshold of hearing, some spatial information discarded (depending on the encoding mode), and written out as a bitstream."
Even though it is wrong, as it ignores temporal masking, one of the key ways in which 'inaudible' information is discarded with MP3 compression. And that the *relationship* between the discarded frequency bands is vital, it's not just a case of doing an FFT and removing any bins with lesser energy.
I guess this is not the place for a technical discussion of lossy codecs, but I live in hope.
Could you elucidate a little on why you think I have no idea about how lossy audio compression works?
Just stating that and talking about PASC does not really tell me anything.
I'm aware it works very differently to a multiband compressor. Some of the artifacts sound quite similar to a badly set one though, and it's the only way I think to communicate what they sound like to me.
The jangling keys test tends to show up problems in audio gear as it contains large amounts of ultrasonic information. Even some standard software equalizers can sound quite bad with it.
The artifacts of MP3 compression, both in the time and frequency domain, are more obvious with uncompressed source material.
:)
:)
Stick a decent mic up in a room, and record some jangling car keys while slowly walking away from the mic till it's almost inaudible, and MP3 will sound awful. (And so will quite a lot of prosumer recording gear recording uncompressed.
To me, MP3 artifacts sound like the distortion you get by putting audio through a multiband compressor with a low ratio and threshold, but with too fast attack and release times. So, you get a constant subtle distortion.
I doubt most people notice, but as I am used to using multibands, I am used to making sure they are working right, and it sounds odd.
Perhaps with heavily audio compressed material, the artifacts of the MP3 compression are less obvious as it's so smashed to start with.
That's a bit like saying that bad food can be improved by removing your sense of taste though.
Thanks for the reply.
I thought it would be a cover or pit band or something. There is nothing wrong with playing covers (it's all about making people enjoy themselves after all).
You don't really need to have recordings or an established fanbase either, as people know pretty much what they are going to get in advance.
It's not the part of the music industry I'm in. I only work with original acts, for better or for worse.
I guess I've been doing it so long I tend to see everything from that perspective.
Out of interest, do they play all original music or are they a cabaret/covers band? And how many people are in the band?
"Before the 20th century musicians made their money by performing. During the 20th century many musicians made their money by recording music. After the advent of the internet musicians will once again make their money by performing and use their recordings as advertising (as everybody but the RIAA bands do now)."
Piffle.
Recorded music is an artform that first appeared in the 20th century, and is distinctly different from acoustic live music. To say people should not be paid for it is to say that only certain musical genres should make money.
Anyway, no one makes any money gigging. You have absolutely no idea how expensive and how time consuming it is. The gigs are to advertise the recordings.
Just work out how much a band with four members and a single roadie/manager/sound guy would have to make every day by gigging to feed everyone and pay the rent.
You'd need to be doing 4-5 gigs a week, every week, all year. How long can a small band do that before they exhaust their fanbase? Probably a couple of months. That's assuming their fanbase is not scattered thinly all over the world, and can actually make it to the local venues the band can reach.
Only the larger or RIAA acts can play places large enough for the economies of scale to kick in and start making profit.
"Even a reference tone of 1KHz gives you only 44 and a bit samples per cycle. That is totally inadequate to describe a complex sound in all its detail, whatever Nyquist says."
Ah, but it is adequate. Any 'complexity' not recorded contains frequencies above half the sample rate. Any sine wave can be described as long as you have more than two samples.
"Anyone, even your grandfather, can hear 5.5KHz. 8 samples per cycle. Better than 4 but hardly accurate. Why don't you try reducing a graph of, say, a stock price over even a day to 8 samples, and then translate it back, without losing information."
If the stock price is low pass filtered in the same way as digital audio before conversion, this is not only possible, but trivial to do correctly.
1 bit SACD works with exactly the same theory as multibit PCM audio. They are no different in principal.
No.
High sample rates do not remove audible stair stepping in the reconstructed waveform.
There is never stair stepping in the reconstructed waveform, regardless of the sampling frequency used, provided both the input and output are correctly band limited to below 1/2 Fs.
It does not matter if you sample at 44.1Khz, 96Khz, 2MHz of 500Khz.
That is the absolute first principle of digital audio recording.
Remember, samples are not signals. What comes out of the reconstruction filter of a D/A converter is not the same as what you see on your computer monitor waveform display.
There is quantization noise, but this is related to bit depth, and is *completely* inaudible due to the use of 1/2 LSB dither. A greater bit depth gives you a lower noise floor, it does not provide "more room for subtleties to be revealed in louder passages".
I don't expect you to understand this. But don't worry, reading the comments on here it appears that very few people do.
The problem is that it's a difficult and non intuitive subject.
People who do not understand the theory basically just keep repeating that Nyquist/Shannon is wrong.
As they do not understand it, no amount of mathematical proof can change their minds.
Arrrgh. The slashcode has removed all the 'less than' signs from my post.
The first and last times I write '1/2 Fs' should have a 'less than' symbol before them.
You don't understand it at all. :(
Nyquist states that there is *exact* reproduction of all frequencies 1/2 Fs.
You only need two and a bit samples to reconstruct a sine and it's phase, and all waveforms can be thought of as a sum of sinewaves of different frequencies and phases.
So, as any sine below 1/2 Fs is reconstructed exactly, so is any possible waveform with harmonics 1/2 Fs.
That won't work.
16/44.1 digital audio is already at the threshold of our perception. Ie, it's had to hear the difference between it and 24/96.
Lossy compression uses a perceptual model based on the way we perceive sound.
So, to create a lossy codec that is of a higher quality than 16/44.1, we need to know how we hear the difference between 16/44.1 and higher sample rate/bit depths. The available data bandwidth can then be concentrated on preserving what is most important to the human ear about that difference.
Since very few people can tell the difference, building a perceptual model of it is very hard.