Another example would be “Channel 2 timecode” files where an interview or production audio sound is recorded on one channel and SMPTE audible timecode is recorded on the other.

Using Joint Stereo in these cases will result in ‘bleed-thru’ of one channel to the other, in particular when using audible SMPTE timecode which ‘sounds like white noise’.

I have written an Ogg Vorbis encoder two years ago for a specific application. In Vorbis the MS stereo implementation is particularly clever, as you do the conversion only after you have already quantized your frequency values. In effect this means that the conversion is lossless by its very definition. Using MS stereo doesn’t change a single sample in audio reproduction, it only affects file size.

My findings using MS stereo were as follows:
1) When compared to mono, using conventional stereo added 100% to the file size, i.e. it doubled it. That was of course what would be expected.
2) When using MS stereo, adding stereo information made the files on average 40%-60% larger instead of 100%. However, there was a big difference depending on the type of source material. All in all, typically two thirds of the resulting file contained M information, and one third S information.

All in all, and as can be calculated from these numbers, using MS stereo on OGG files helped make files on average 25% smaller as compared to discrete stereo, with _exactly_ the same decoded result. Because of the losslessness of MS stereo in OGG, the OGG encoders don’t even give you the option of choosing whether you use MS stereo or not. The only reason I know this is that I researched the matter when creating my own encoder.

MP3 files are slightly different because MS stereo is handled at a different stage of compression. Due to this the results are not bit-accurate, but for the same file size, and using a proper encoder, MS stereo still will give better results.

First of all, there has still not been a single double-blind test ever done that proves that anyone, ANY-ONE, could hear the difference between direct sound and sound properly sampled at 44.1 kHz 16 bits with high quality A/D + D/A converters. This has been the case since the early 1980′s, and if the proof hasn’t surfaced in 30 years, I don’t think it ever will. Repeating digital inferiority without a shred of proof just doesn’t make it true.

Second, turning normal Left-Right stereo into Mid-Side stereo and back is a lossless operation. You site FM radio as an example of how MS stereo is somehow technically worse than mono. This is quite an incorrect analogy. The reason why FM stereo radio has more noise than FM mono radio is because of some unfortunate technical decisions which made stereo FM receivers less expensive to make, but which at the same time limited the signal-to-noise ratio of the side information channel by about 20 dB. The noise is FM stereo is an artifact specific to FM radio and has nothing to do with digital MS stereo.

MS stereo in digital is just based on the fact that if you losslessly convert your LR stereo to MS representation (M = L+R, S = L-R), numbers in the S channel tend to become much smaller, both in the time and frequency domains. Smaller numbers are easier to compress, so the MS representation allows for more efficient compression. And compression, after all, is the whole point of MP3, AAC and OGG (all of which can use MS stereo in one form or another). E.g. MP3 can select on a frame-to-frame basis whether to use MS or LR stereo. If the left and right channels correlate little enough that MS stereo wouldn’t help, the frame can be encoded as LR stereo. When properly done, there is absolutelu _no_ spatial image lost. Vice versa, if the bitrate is set to a certain limit, this will yield a better-sounding file, also in the spatial domain.

As for recordings made with phase errors between the left and right channels, there are no problems in representing them in the MS stereo format. Their representation will just be less efficient because the S signal doesn’t get as small as with a properly aligned recording. As a result you’d need a larger file for the same quality.

Naturally there are bad implementation, but e.g. current versions of LAME are quite good in selecting the optimal compression mode automatically when using MS stereo.

I’d like to also comment the original article’s definition of intensity stereo. In intensity stereo part of the audio spectrum is indeed encoded only as mono, but what is missing from the article is that then panning information is sent on a critical band basis for that part of the spectrum. So, e.g. if you encode intensity stereo and you have two sine tones, e.g. at 5 kHz in the left channel and 11 kHz in the right channel, that will be encoded as one mono signal that has both the frequencies. Frequency band panning information however will make it possible to represent it properly so that the listener will indeed hear the 5 kHz tone in his left and 11 kHz tone in his right loudspeaker. However, because the frequency bands are quite wide, intensity stereo cannot represent properly a situation where there is a 10 kHz signal in the left channel and 10.1 kHz signal in the right channel. In this case the bouth sine waves would be, depending on the encoder, likely just be played in the center. That’s the price you pay for intensity stereo: you can have a high-hat in the right channel, but brass instrument higher overtones cannot be represented in the left channel at the same time. Nevertheless, when encoding at very small bitrates, this smudging of the stereo image is usually less bothersome than other artifacts like warbling, so it is a very efficient way of audio compression.

L-> 16, R->8

Joint stereo encodes as the sum and difference of the two channels. In this case, the sum is 24 and the difference is 8.

S: 24 D: 8

You can find the original Left and Right values with these formulae: L = S – D and R = S – L.

L=24-8
L=16

R=24-L
R=24-(16)
R=8

Since we got back the original values exactly, and since we can do this regardless of what the values are (I dare you to try and find a case where this doesn’t work), joint stereo is lossless.

However, the computer must allocate enough space to store both values exactly or all that losslessness would be in vain. Assuming we are dealing with 16 bits per channel per sample integer audio (the most common type), to do this in all cases (including the worst case) we would need 17 bits for the sum channel and 16 bits for the difference channel. This means that instead of needing 32 bits per sample (2 channels*16) we would need 33. This means we would actually need MORE space to store joint stereo than plain stereo. On top of that, computers store bits in groups called bytes, and since bytes are almost always eight bits long, we would need 40 bits to store a sample of joint-stereo audio, since under most circumstances, computers don’t store bits by themselves (only in bytes).

However, this is where the magic of data compression kicks in. I’m not going to go into details, but a well implemented data compression algoritm can compress the joint-stereo data much better than individual channels, assuming there is much in common with the left and right channels. This is how FLAC works, even though it calls joint stereo “inter-channel decorrelation” and says joint stereo is lossy when they are really the same (They are probably trying to disassociate joint stereo with the additional sound degradation it can cause when using it with MP3 and other lossy compression algorithms).

The problem is that lossless compression can’t shrink audio enough for many purposes, like sharing music online (or sharing videos online with sound in them). This is where technologies like MP3 come in, though MP3 is a rather outdated technology (Vorbis is better, spec-wise at least). These are lossy compression algorithms, and they can also usually compress joint stereo better (by cutting out the less important data, especially in the difference channel), especially at lower bitrates. The problem here is that the lossiness can be more noticable when using joint stereo, especially at very low and high bitrates (as the article and other comments describe).

So if you can find the disk space/bandwidth to compress losslessy, you can (and should provided the left and right channels are similar; this is usually the case) go joint-stereo with no extra degradation.