## Emphasizing a Presumed Difference between OGG and MP3 Sound Compression

In this posting from some time ago, I wrote down certain details I had learned about MP3 sound compression. I suppose that while I did write, that the Discreet Cosine Transform coefficients get scaled, I may have missed to mention in that same posting, that they also get quantized. But I did imply it, and I also made up for the omission in this posting.

But one subject which I did mention over several postings, was my own disagreement with the practice, of culling frequency-coefficients which are deemed inaudible, thus setting those to zero, just to reduce the bit-rate in one step, hoping to get better results, ‘because a lower initial bit-rate also means that the user can select a higher final bit-rate…’

In fact, I think that some technical observers have confused two separate processes that take place in MP3:

1. An audibility threshold is determined, so that coefficients which are lower than that are set to zero.
2. The non-zero coefficients are quantized, in such a way that the highest of them fits inside a fixed maximum, quantized value. Since a scale-factor is computed for one frequency sub-band, this also implies that close to strong frequency coefficients, weaker ones are just quantized more.

In principle, concept (1) above disagrees with me, while concept (2) seems perfectly fine.

And so based on that I also need to emphasize, that with MP3, first a Fast-Fourier Transform is computed, the exact implementation of which is not critical for the correct playback of the stream, but the only purpose of which is to determine audibility thresholds for the DCT transform coefficients, the frequency-sub-bands of which must fit the standard exactly, since the DCT is actually used to compress the sound, and then to play it back.

This FFT can serve a second purpose in Stereo. Since this transform is assumed to produce complex numbers – unlike the DCT – it is possible to determine whether the Left-Minus-Right channel correlates positively or negatively with the Left-Plus-Right channel, regarding their phase. The way to do this effectively, is to compute the dot-product between two complex numbers, and to see whether this dot-product is positive or negative. The imaginary component of one of the sources needs to be inverted for that to work.

But then negative or positive correlation can be recorded once for each sub-band of the DCT as one bit. This will tell, whether a positive difference-signal, is positive when the left channel is more so, or positive if the right channel is more so.

You see, in addition to the need to store this information, potentially with each coefficient, there is the need to measure this information somehow first.

But an alternative approach is possible, in which no initial FFT is computed, but in which only the DCT is computed, once for each Stereo channel. This might even have been done, to reduce the required coding effort. And in that case, the DCT would need to be computed for each channel separately, before a later encoding stage decides to store the sum and the difference for each coefficient. In that case, it is not possible first to determine, whether the time-domain streams correlate positively or negatively.

This would also imply, that close to strong frequency-components, the weaker ones are only quantized more, not culled.

So, partially because of what I read, and partially because of my own idea of how I might do things, I am hoping that OGG sound compression takes this latter approach.

Dirk

## Scale Factor == Step Size

It could be the case, that in my own postings I referred to a ‘Scale Factor’, which would come prior to Quantization. In other works of reference, the term ‘Quantization Step’ might appear. As far as I am concerned, these terms are synonymous.

The goal could be to start with a maximum value as input, and to find a way to quantize it and all lower values, to arrive at a maximum quantized integer value. One would divide the (absolute) first by the second input value, to find this parameter, for an interval of time.

Dividing all the sample-values in the interval by the resulting parameter, will yield the maximum, quantized integer. And this parameter will also be equal to the minimum difference between the sample values, leading to two different quantized integers.

## aptX and Delta-Modulation

I am an old-timer. And one of the tricks which once existed in Computing, to compress the amount of memory that would be needed, just to store digitized sound, was called “Delta Modulation”. At that time, the only ‘normal’ way to digitize sound was what is now called PCM, which often took up too much memory.

And so a scheme was devised very early, by which only the difference between two consecutive samples would actually stored. Today, this is called ‘DPCM‘. And yet, this method has an obvious, severe drawback. If the signal contains substantial amplitudes, associated with frequencies that are half the Nyquist Frequency or higher, this method will clip that content, and produce dull, altered sound.

Well one welcoming fact which I have learned, is that this limitation has essentially been overcome. One commercial domain in which this has been overcome, is with the compression scheme / CODEC named “aptX“. This is a proprietary scheme, owned by Qualcomm, but is frequently used, as the chips manufactured and designed by Qualcomm are installed into many devices and circuits. One important place this gets used, is with the type of Bluetooth headset, that now has high-quality sound.

What happens in aptX, requires that the band of frequencies which start out as a PCM stream, needs to get ‘beaten down’ into 4 sub-bands, using a type of filter known as a “Quadrature Mirror Filter“. This happens in two stages. I know of a kind of Quadrature Mirror Filter which was possible in the old analog days, but have had problems until now, imagining how somebody might implement one using algorithms.

The analog approach required, a local sine-wave, a phase-shifted local sine-wave, a balanced demodulator used twice, and a phase-shifter which was capable of phase-shifting a (wide) band of frequencies, without altering their relative amplitudes. This latter feat is a little difficult to accomplish with simple algorithms, and when accomplished, typically involves high latency. aptX is a CODEC with low latency.

The main thing to understand about a Quadrature Mirror Filter, implemented using algorithms in digital signal processing today, is that the hypothetical example the WiKi article above cites, using a Haar Wavelet for H0 and its complementary series for H1, actually fails to implement a quadrature-split in a pure way, and was offered just as a hypothetical example. The idea that H1( H0(z) ) always equals zero, simply suggested that the frequencies passed by these two filters are mutually exclusive, so that in an abstract way, they pass the requirements. After the signal is passed through H0 and H1 in parallel, the output of each is reduced to half the sampling rate of the input.

What Qualcomm explicitly does, is to define a series H0 and a series H1, such that they apply “64 coefficients”, so that they may achieve a frequency-split accurately. And it is not clear from the article, whether the number of coefficients for each filter is 64, or whether their sum for two filters is 64, or the sum of all six. Either way, this implies a lot of coefficients, which is why dedicated hardware is needed today, to implement aptX, and this dedicated hardware belongs to the kind, which needs to run its own microprogram.

Back in the early days of Computing, programmers would actually use the Haar Wavelet, because of its computational simplicity, even though doing so did not split the spectrum cleanly. And then this wavelet would define the ‘upper sideband’ in a notional way, while its complementary filter would define the notional, ‘lower sideband’, when splitting.

But then the result of this becomes 4 channels in the case of aptX, each of which has 1/4 the sampling rate of the original audio. And then it is possible, in effect, to delta-modulate each of these channels separately. The higher frequencies have then been beaten down to lower frequencies…

But there is a catch. In reality, aptX needs to use ‘ADPCM‘ and not ‘DPCM’, because it can happen in any case, that the amplitudes of upper-frequency bands could be high. ADPCM is a scheme, by which the maximum short-term differential is computed for some time-interval, which is allowed to be a frame of samples, and where a simple division is used to compute a scale factor, by which these differentials are to be quantized.

This is a special situation, in which the sound is quantized in the time-domain, rather than being quantized in the frequency-domain. Quantizing the higher-frequency sub-bands has the effect of adding background – ‘white’ – noise to the decoded signal, thus making the scheme lossy. Yet, because the ADPCM stages are adaptive, the degree of quantization keeps the level of this background noise at a certain fraction, of the amplitude of the intended signal.

And so it would seem, that even old tricks which once existed in Computing, such as delta modulation, have not gone to waste, and have been transformed into something more HQ today.

I think that one observation to add would be, that this approach makes most sense, if the number of output samples of each instance of H0 is half as many, as the number of input samples, and if the same can be said for H1.

And another observation would be, that this approach does not invert the lower sideband, the way real quadrature demodulation would. Instead, it would seem that H0 inverts the upper sideband.

If the intent of down-sampling is to act as a 2:1 low-pass filter, then it remains productive to add successive pairs of samples. Yet, this could just as easily be the definition of H1.

Dirk

(Edit 06/20/2016 : ) There is an observation to add about wavelets. The Haar Wavelet is the simplest kind:


H0 = [ +1, -1 ]
H1 = [ +1, +1 ]


And this one guarantees that the original signal can be reconstructed from two down-sampled sub-bands. But, if we remove one of the sub-bands completely, this one results in weird spectral results. This can also be a problem if the sub-bands are modified in ways that do not match.

It is possible to define complementary Wavelets, that are also orthogonal, but which again, result in weird spectral results.

The task of defining ones, which are both orthogonal and spectrally neutral, has been solved better by the Daubechies series of Wavelets. However, the series of coefficients used there are non-intuitive, and were also beyond my personal ability to figure out spontaneously.

The idea is that there exists a “scaling function”, which also results in the low-pass filter H1. And then, if we reverse the order of coefficients and negate every second one, we get the high-pass filter H0, which is really a band-pass filter.

To my surprise, the Daubechies Wavelets achieve ‘good results’, even with a low number of coefficients such as maybe 4? But for very good audio results, a longer series of coefficients would still be needed.

One aspect to this which is not mentioned elsewhere, is that while a Daubechies Wavelet-set could be used for encoding, that has a high order of approximation, it could still be that simple appliances will use the Haar Wavelet for decoding. This could be disappointing, but I guess that when decoding, the damage done in this way will be less severe than when encoding.

The most correct thing to do, would be to use the Daubechies Wavelets again for decoding, and the mere time-delays that result from their use, still fall within the customary definitions today, of “low-latency solutions”. If we needed a Sinc Filter, using it may no longer be considered so, and if we needed to find a Fourier Transform of granules of sound, only to invert it again later, it would certainly not be considered low-latency anymore.

And, when the subject is image decomposition or compression, it is a 2-dimensional application, and the reuse of the Haar Wavelet is more common.

## Why the Temporal Resolution of MP3s is Poor.

I have spent a lot of my private time, thinking about lossy sound compression, and then, simplifying my ideas to something more likely to have been implemented in actual MP3 compression. In order to understand this concept, one needs to be familiar with the concept of Fourier Transforms. There are two varieties of them, which are important in sound compression: The “Discreet Fourier Transform” (‘DFT’), and the “Discreet Cosine Transform” (‘DCT’), the latter of which has several types again.

I did notice that the temporal resolution of MP3s I listen to is poor, and it was an important realization I finally came to, that this was not due to the actual length of the sampling window.

If we were to assume for the moment that the sampling interval was 1024 samples long – and for MP3, it is not – then to compute the DFT of that would produce 1024 frequency coefficients, from (0-1023 / 2) cycles / sampling interval. Each of these coefficients would be a complex number, and the whole set of them can be used to reconstruct the original sample-set accurately, by inverting the DFT. The inverse of the DFT is actually the DFT computation again, but with the imaginary (sine) component inverted (negated).

But, MP3s do not use the DFT, instead using the DCT, the main difference in which is, that the DCT does not record a complex number for each coefficient, rather just stating a real number, which would normally correspond only to the cosine function within the DFT… Admittedly, each of these absolute amplitudes may possibly be negated.

If the time-domain signal consisted of a 5 kHz wave, which was pulsating on and off 200 times per second – which would actually sound like buzzing to human ears – then the DCT would record a frequency component at 5kHz, but as long as they are not suppressed due to the psychoacoustic models used, would also record ‘sidebands’ at 4800 and at 5200 Hz, each of which has 1/2 the amplitude of the center frequency at 5 kHz. I know this, because for the center frequency to be turned on and off, it must be amplitude modulated, virtually. And so what has this time-domain representation, even though this happens faster than once per sampling window, also has a valid frequency-domain representation.

When this gets decoded, the coefficient-set will reproduce a sample-set, whose 5 kHz center frequency again seems to ‘buzz’ 200 times per second, due to the individual frequency components interfering constructively and then destructively, even though they are being applied equally across the entire sampling interval.

But because the coefficient-set was produced by a DCT, it has no accurate phase information. And so the exact time each 5 kHz burst has its maximum amplitude, will not correspond to the exact time it did before. This will only seem to correct itself once per frame. If the sampling interval was truly 1024 samples long, then a frame will recur every 512 samples, which is ~80 times per second.

Now the question could be asked, why it should not be possible to base lossy audio compression on the DFT instead. And the answer is that in principle, it would be possible to do so. Only, if each coefficient-set consisted of complex numbers, it would also become more difficult to compress the number of kbps kept in the stream, in an effective way. It would probably still not be possible to preserve the phase information perfectly.

And then as a side-note, this one hypothetical sample-set started out as consisting of real numbers. But with the DFT, the sample-set could carry complex numbers as easily as the coefficient-set did. If the coefficients were compressed-and-simplified, then the samples reproduced would probably end up being so, with complex values. In a case like this, the correct thing to do is to ignore the imaginary component, and only output the real component, as the decoded result…

When using a DCT to encode a stream of sound, which is supposed to be continuous, a vulgarization of the problem could be, that the stream contains ‘a sine wave instead of a cosine wave’, which would therefore get missed by all the sampling intervals, because only the product with the cosine function is being computed each time, for a specific coefficient. The solution that comes from the Math of the DCT itself is, that the phase of the unit vector generally rotates 90 degrees ~from each frame to the next~. To the best of my understanding, two sampling intervals will generally overlap by 50% in time, resulting in one frame half as long. It may be that the designers only compute the odd-numbered coefficients. Then, the same coefficient belonging to the next frame should be aware of this wave notwithstanding. Further, the sampling intervals are made to overlap when the stream is decoded again, such that a continuous wave can be reconstructed. ( :1 )

The only question I remain curious about, is whether a need exists when encoding with a DCT, to blend any given coefficient as belonging to two consecutive frames, the current one plus the previous one.

While it can be done, to use rectangular sampling windows for encoding, the results from that are likely to be offensive to listen to. So in practice, Blackman Windows should ideally be used for encoding (that are twice as long as a frame).

The choice of whether decoders should use a Hanning Window or a Linear Taper, can depend on what sort of situation should best be reproduced.

Decoding with a linear taper, will cause crescendos to seem maximally smooth, and perfectly so if the crescendo is linear. But considering that linear crescendos might be rare in real music, a Hanning Window will minimize the distortion that is generated, when a burst of sound is decoded, just as a Blackman Window was supposed to do when encoding. Only, a Blackman Window cannot be used to decode, because coefficients being constant from one frame to the next would result in non-constant (output) sample amplitudes.

Dirk

(Edit 05/18/2016 : ) One related fact should be acknowledged. The DCT can be used to reconstruct a phase-correct sample-set, if non-zero even-numbered as well as odd-numbered coefficients are included. This follows directly from the fact that a ‘Type 3′ DCT is the inverse of the ‘Type 2′. But, the compression method used by several codecs is such, that a psychoacoustic model suppresses coefficients, on the assumption that they should be inaudible, because they are too close spectrally, to stronger ones. This would almost certainly go into effect, between complementary even-numbered and odd-numbered DCT coefficients.

( 05/31/2016 : ) One detail which was not made clear to me, was whether instead, coefficients which are in the same sub-band as one that has a stronger peak, are merely quantized more, due to the scale-factor of that sub-band being higher, to capture this higher peak. This would strike me as favorable, but also results in greater bit-rates, than what would follow, from setting supposedly-inaudible coefficients to zero. Due to Huffman Encoding, the bit-length of a more-quantized coefficient, is still longer than that for the value (zero).

In any type of Fourier Transform, signal energy at one frequency cannot be separated fully from energy measured at a frequency different by only half a cycle per frame. When the difference is at least by one cycle per frame, energy and therefore amplitude become isolated. This does not mean however, that the presence of a number of coefficients equal to the number of samples, is always redundant.

And so, one good way to achieve some phase-correctness might be, to try designing a codec, which does not rely too strongly on the customary psychoacoustic models. For example, a hypothetical codec might rely on Quantization, followed by Exponential-Golomb Coding of the coefficients, being sure to state the scale of quantization in the header information of each frame.

It is understood that such approaches will produce ‘poorer results’ at a given bit-rate. But then, simply choosing a higher bit-rate (than what might be appropriate for an MP3) could result in better sound.

And then, just not to make our hypothetical codec too primitive, one could subdivide the audible spectrum into 8 bands, each one octave higher than the previous, starting from coefficient (8), so that each of these bands can be quantized by a different scale, according to the Threshold Of Audibility. These Human Loudness Perception Curves may be a simple form of psychoacoustics, but are also thought to be reliable fact, as perceived loudness does not usually correspond consistently to uniform spectral distribution of energy.

Parts of the spectrum could be quantized less, for which ‘the lower threshold of hearing’ is lower with respect to a calculable loudness value, at which the Human ears are thought to be uniformly sensitive to all frequencies.

Assigning such a header-field 8 times for each frame would not be prohibitive.

1: ) ( 05/31/2016 ) An alternative approach, which the designers of MP3 could just as easily have used, would have been first to compute the DCT, including both even- and odd-numbered coefficients F(k), but then to derive only the even-numbered coefficients from that. The best way would have been, for even numbered, derived coefficient G(k) to be found as

r = F(k)

i = F(k+1) – F(k-1)

G(0) = F(0)

G(k) = sqrt( r^2 + i^2 )