|
Audio Asylum Thread Printer Get a view of an entire thread on one page |
For Sale Ads |
In Reply to: Is MLP flawed? posted by uzun on April 03, 2003 at 09:41:11:
I suggest doubters read this excerpt from "The MLP lossless compression system" by Gerzon, Craven, Stuart — published at the 17th AES (sorry -- I cannot include illustrations):-/ . . . . . . . . . . . . . . . . . . . . . .
4.7 Buffering
We have explained that while normal audio signals can
be well predicted, there will be occasional fragments
like sibilants, synthesised noise or percussive events that
have high entropy.MLP uses a particular form of stream buffering that can
reduce the variations in transmitted data rate, absorbing
transients that are hard to compress.FIFO memory buffers are used in the encoder and
decoder as shown in Figure 13. These buffers are
configured to give a constant notional delay across
encode and decode. This overall delay is small –
typically of the order of 75ms. To allow rapid start-up or
cueing, the FIFO management minimises the part of the
delay due to the decoder buffer. So, this buffer is
normally empty and fills only ahead of sections with
high instantaneous data rate.During these sections, the decoder’s buffer empties and
is thus able to deliver data to the decoder core at a
higher rate than the transmission channel is able to
provide. In the context of a disc, this strategy has the
effect of moving excess data away from the stress peaks,
to a preceding quieter passage.The encoder can use the buffering for a number of
purposes, e.g.:~ Keeping the data-rate below a preset (format) limit.
~ Minimising the peak data rate over an encoded
section.Figure 14 shows an example of the latter. The entropy-coded
data rate from the encoder core is shown along
with the buffered result. The buffered data has a
characteristic flat-topped curve. This is not due to
clipping or overload, but to rate absorption in the
encoder/decoder FIFOs.Another illustration of data-rate minimisation is shown
in Figures 15 and 16. Again the encoded data rate is
plotted through a 30-second 96kHz 24-bit 6-channel
excerpt featuring a close recording of a jazz saxophone.
Figure 15 indicates the underlying compression with the
encoder set to limit above 9·5Mbps. The minimum-rate
encode shown in Figure 16 makes long-term (but low
occupancy) use the decoder buffer.It should be obvious that the situation in Figure 16 is
preferable if the transmission channel (maybe DVD
disc) has other calls on the bandwidth – for example
bandwidth to transmit associated picture or text.
Figure 17 shows how hard-to-compress signals can be
squeezed below a preset format limit. This 30-second
96kHz 24-bit recording features closely recorded
cymbals in 6 channels. At the crescendo this signal is
virtually random and the underlying compressed data
rate is 12·03Mbps. Buffering allows the MLP encoder to
hold the transmitted data rate below 9·2Mbps by filling
the decoder buffer to a short-term maximum of 86kbyte
(bottom curve).Figure 18 shows the potential for peak data-rate
reduction on this item with different amounts of
available FIFO memory.
Follow Ups:
It is flawed, since MLP did not meet their own objectives, and had to be "patched up" to make it work.This is undeniable, the amount of importance you place on this hack however is subjective. What we can say with 100% certainty is that your not getting the "master tape" sound, with watermarking and HF roll off applied.
No matter how good your DA's get, you can never get back what has been ruined.
> What we can say with 100% certainty is that your not getting the "master tape" sound, with watermarking and HF roll off applied. <Watermarking can and is applied to SACD, but so is a mandatory 50kHz filter which removes all content above that point because of the huge amount of noise present in the DSD stream, unlike the flat response of MLP out to 96kHz.
If you think you're getting "master tape" sound with SACD, then you've got another thing coming buster! ;-)
or at least the use of some sort of buffer, in any sort of streaming where the rate varies.The fact that MLP uses a FIFO is not surprising. What Dave P. seems to be implying that they have not used a large enough FIFO, so that they still have problems.
If the offending signal is sustained the problem will occur slightly later.If a signal is random and has high level high frequency content the information cannot be packed below the maximum allowable range.
With normal music it will practically never occur because music is by nature rather 'predictable'. And if it happens the measures are simple and have no significant or detectable impact on sonics.
I similar problem exist with dsd where they strongly suggest to keep the signal input -6dB below the maximum range of the ad converter.
With good recording (for dsd) and mastering practises these 'overload' problems are non issues.
This post is made possible by the generous support of people like you and our sponsors: