Marantz SA-11S3 Quick reference guide

Introduction
Since the start of the era of digital recording of music, technology
has been faced with the challenge of maximising the accuracy
of reproduction. The initial problem was the small number of bits
available for the conversion of analogue signals into digital. In the
early ‘80s the first attempts were made to create more audible
resolution than the bits actually produced. In this way, the first
CD player from Philips and Marantz worked with the 14-Bit-D/A-
Converter using a process called noise-shaping and achieved real
precision of 16 bits and four-fold oversampling, and thus played a
leading role.
In recent years, the issue has in fact been reversed, because cur-
rent master and media can easily deal with 24 or even 32-bit and
provide the necessary digital filtering for playing up to 48 bits per
sample. Now the issue centers on reducing the surplus resulting
resolution and the ensuing huge bandwidths by hundreds of
kilohertz as efficiently as possible to usable frequencies and the
common resolutions of 16 (CD) to 24 bits. If the excess bits were
shortened to simply remove the excess bits, audible errors would be
created and any higher resolution gained would be irretrievably lost.
In the SA-11S3 Marantz has implemented new algorithms for digital
signal processing developed in-house for the first time and the result
is a surprisingly low loss of resolution. For this purpose, Marantz
has utilised new signal processing technology previously exclusively
reserved for high-end, professionally equipped mastering studios.
Solution
To maximise audible resolution, three methods are combined:
Oversampling, Noise Shaping and Dithering.
Through oversampling intermediate samples between 2 original
samples are calculated, thus achieving a multiplication of the
sampling rate. To calculate the intermediate samples with
sufficient accuracy the resolution of the audio data needs to be
increased significantly. In our application we found additional
24 bits adequate giving us a total resolution of 48 bits.
The noise shaping offsets digital noise components through clever
filtering from the audible frequency range into the frequency range
beyond 20 kHz, which is no longer perceptible.
With dithering, the desired signal is modulated with noise to
reduce inaccuracies caused through the forcible conversion of reso-
lutions. The type and distribution of the noise determine the quality
of the sonic result. Marantz established the optimal variant through
extensive listening tests and trial series.
As a result of the correct combination of the three methods and the
correct choice of the respective algorithm and its application, more
details than the simple conversion of the original signal could be
perceived, thus significantly higher-resolution sound reproduction
could be achieved.
www.marantz.eu Marantz Musical Mastering | SA-11S3 | 1
White paper
Marantz Musical Mastering
SA-11S3

Oversampling
Unlike many procedures used previously, Marantz deploys no rational
numbers for the upsampling of the sampling rate, e.g. 44.1 kHz from
a CD to 96kHz (factor 2.1768 ...), as this is often results in a deteriora-
tion of the “flow” and the “rhythm” of the playback. In its frequency
multiplier Marantz uses exclusively integer factors, and converts the
data with a FIR filter (finite impulse response filter) at either to 352.8
or 384 kilohertz. At the same time, this increases the quantisation
depth up to 48 bits; depending on the corresponding resolution of
the input signal (16 to 24 bits) results in a 2 to 8-fold oversampling.
Noise Shaping
The noise shaping quantifies a mathematical return of rounding
errors back to the beginning of the processing of the audio data.
The effect of this feedback is similar to the “transfer” in manual,
column-by-column addition to or subtraction from two numbers on
the paper. This feedback from the rounding error of the quantiser
is offset against the next digital word (1 sample with up to 48 bits),
added up along with the next random value of the dither and in turn
sent to the quantiser. Over time each rounding error of the voltage
axis (bits) shifts to a correct value on the time axis (sampling
frequency) in this way. The result corresponds more accurately with
the original audio signal with the higher resolution and therefore
sounds better than simply rounding or truncation the signal. The
random element of the dithering signal also prevents the occur-
rence of noise by the feedback of the transfer to the input, as
could occur with the previously mentioned noise shaper procedure
without dithering.
The noise shaper filter only acts with the LSB (the least significant
bit, the “smallest” bit) and returns it to the input of the quantizer
Dithering
Dithering is, put very simply, the signal mixed with noise. In the
reduction from high to low resolution, artifacts which can be
perceived as a distortion occur, as the patterns now appear that do
not belong to an actual music signal.
This can be illustrated through graphic sequencing. The colour
gradient of the diagram, originally 24-bit, was reduced to 4 bits to
illustrate the effect. The problem is transferable between graphics
and audio.
Here is a diagram quantised with 24-bit
colour depth. In this analogy it corre-
sponds with an original, high-resolution
audio signal. It consists of a similar
representation of an analogue image,
with smooth, seamless colours.
If the graphics to illustrate a lower colour
depth are added below, anomalies that
do not match the original (24-bit) appear.
Even artifacts, new image details will be-
come visible in the form of ring patterns
that do not belong to the original image.
The same occurs in the downsampling
of audio signals. When flattening higher
resolutions, interference (distortion)
is created that does not belong to the
original.
If the reduced image is allocated to a
suitable noise, artifacts occurring as a
result of the reduction of the original
low-resolution image disappear almost
completely. With the displaced dither, the
image has a significantly greater resem-
blance to the the original high resolution,
despite the lower depth in colour.
Something similar takes place In the
audio processing. With the right dither
displacement, the rounding error is dis-
tributed in a statistically favourable man-
ner, the sound is closer to the original.
www.marantz.eu
White paper
Marantz Musical Mastering | SA-11S3 | 2



Original:
24-Bit graphics
Linear scaling:
Reduction to 4 bits.
Quantiser
Noise Shaper Filter
24 Bit
MSB
24 Bit
Dither
48 Bit
Music
Data
24 Bit
LSB

To ensure the generated random signal (noise) remains random
from an acoustic point of view, it may not be repeated at audible
repetition rates. To achieve a minimum repetition rate of 1 second
or higher very long cycle of 180,000 samples are needed. A random
number generator was specifically developed for this purpose.
When development was in the initial stage there was a noise
signal with a homogeneous uniform distribution of frequency and
amplitude, a so-called ‘white noise’. The graph below shows the
scaled amplitude statistics of this noise signal.
As can be seen the 180.000 samples show very equal probability
between 0 and 1. This means that all amplitudes between 0 and 1
occur equally often.
Listening tests and research have shown that this linear probability
function can be improved further if 2 such noise generators are
used in parallel. Below the scaled amplitude statistics of such a
noise generator.
So far we investigated only the amplitude behaviour of the noise
generator. The frequency behaviour (e.g. spectrum) of our noise
generator looks like this.
It can be seen that all frequencies have about the same amplitude –
thus our noise generator really generates ‘white noise’.
www.marantz.eu
White paper
Marantz Musical Mastering | SA-11S3 | 3
White noise: a very uniformly distributed random signal.
The resulting noise retains that perfect,yet weighted, random distribution



Again several listing tests have been conducted and finally the best
sonic performance is achieved with filtered white noise. A high-pass
filter led to a significant improvement in sound. Below the plot of
the spectrum of the final dither signal.
This improves the noise spectrum of the resulting
dithered audio signal
Here you can see the distribution of the dither signal, along with its
increasing frequency weighted influence on the audio spectrum
The weighted noise of the dither signal developed by Marantz is
now stronger at high frequencies than at low frequencies and po-
tential interference components are shifted to the frequency range
that is no longer perceptible, beyond the auditory threshold.
Result
This advanced method of signal processing increases
and maintains the high resolution of both new and old
recordings. The distinct improvement in sound is
manifested in detail in individual instruments, space and
richer tones.
www.marantz.eu
White paper
Marantz Musical Mastering | SA-11S3 | 4