Spatialization Parameter Estimation in MDCT Domain for Stereo Audio

K. Suresh & Akhil Raj R.
Signal Processing: An International Journal (SPIJ), Volume (9) : Issue (5) : 2015 66
Spatialization Parameter Estimation in MDCT Domain for
Stereo Audio
K. Suresh sureshk@cet.ac.in
Department of Electronics & Communication
Government Engineering College
Wayanad, Kerala, India, 670644
Akhil Raj R. akhilraj.89@gmail.com
Department of Electronics & Communication
College of Engineering, Thiruvananthapuram
Kerala, India, 695016
Abstract
For representing multi-channel audio at low bit rate parametric coding techniques are used in
many audio coding standards. An MDCT domain parametric stereo coding algorithm which
represents the stereo channels as the linear combination of the ‘sum’ channel derived from the
stereo channels and a reverberated channel generated from the ‘sum ’channel has been reported
in literature. This model is inefficient in capturing the stereo image since only four parameters per
sub-band is used as spatialization parameters. In this work we improve this MDCT domain
parametric coder with an augmented parameter extraction scheme using an additional
reverberated channel. We further modify the scheme by using orthogonalized de-correlated
channels for analysis and synthesis of parametric stereo. A synthesis scheme with perceptually
scaled parameter set is also introduced. Finally we present, subjective evaluation of the different
parametric stereo schemes using MUSHRA test and the increased the perceptual audio quality of
the synthesized signals are evident from these test results.
Keywords: Parametric Audio Coding, MDCT, Parametric Stereo.
1. INTRODUCTION
For multichannel audio compression with reasonable quality at low bit-rates, parametric coding
has emerged as a suitable method with many potential applications [1,4,11]. In multichannel
audio, significant amount of inter channel redundancies present along with perceptual
irrelevancies and statistical redundancies. Effective removal of the inter channel redundancy and
perceptual irrelevancy is required for low bit rate compression of multichannel audio. Considering
the constituent channel data individually, we can apply mono audio compression methods to
remove perceptual irrelevancies and statistical redundancies. To remove inter channel
redundancies, cross channel prediction based methods were suggested [2]. However, in most of
the cases, the correlation between the channels is low and the cross channel prediction based
methods will not result in significant compression [3]. Another strategy for suppressing the inter
channel redundancy is through parametric coding. In parametric coding, the encoded data
consists of a mono ‘sum’ channel derived from the individual channels and model parameters
representing the spatialization cues. Binaural cue coding (BCC) introduced in [8,9,10] and
parametric stereo coding method introduced in [11] are examples for parametric multichannel
audio coding. BCC uses inter channel level difference (ICLD), inter channel time difference
(ICTD) and inter channel coherence (ICC) as parameters for spatial audio modeling. In the
encoder, a ‘sum’ channel is derived by adding the individual channels followed by energy
equalization. The ‘sum’ channel is compressed using any of the existing audio coding algorithms.
The spatialization parameters are extracted in a frame by frame basis and quantized for compact
representation. In the decoder, the multichannel audio is synthesized using the ‘sum’ channel

signal and the spatialization parameters; to synthesis the required level of correlated side
channels, decorrelated signals are generated from the ‘sum’ channel. In the human auditory
system, the processing of binaural cues is performed on a non-uniform frequency scale [12,13].
Hence, in order to estimate the spatial parameters from the given input signal, it is desirable to
transform its time domain representations to a representation that resembles this non uniform
frequency scale. This transformation is achieved either by using a hybrid Quadrature Mirror Filter
(QMF) bank or by grouping a number of bands of a uniform transform such as DFT [9,11].
However, in practice, for audio coding purpose, spectral representations such as Modified
Discrete Cosine Transform (MDCT) are used which has the advantage of time domain alias
cancellation and better energy compaction. Hence additional filter bank analysis or transform is
needed for the parameter extraction in the encoder and for the synthesis in the decoder. The
spatialization parameter extraction and ‘sum’ channel formation is done as a pre-processing step
in the encoder; conversely, the stereo synthesis is a post-processing step in the decoder.
Similarly, the de-correlated channel generation in the decoder is done either by time domain
convolution or equivalent DFT domain multiplication [9,11]. MDCT domain analysis and synthesis
of reverberation for parametric coding of stereo audio has been proposed in [14]. Spatialization
parameter extraction and stereo synthesis from the ‘sum’ channel are done in the MDCT domain.
For parameter estimation, the MDCT coefficients are divided in to twenty two non-uniform blocks
and an analysis by synthesis scheme in the MDCT domain is used. The stereo channels are
approximated to the linear combination of the ‘sum’ channel and a reverberated channel derived
from the ‘sum’ channel. Four parameters are extracted from each block and encoded as the side
information. The spatialization parameters such as ICTD, ICLD and ICC are not estimated
directly. Instead, the de-correlated channel used for stereo synthesis in the decoder is generated
in the encoder and it is used to estimate the synthesis coefficients through least square
approximation method. The parametric coder realized using this parameter extraction method is
capable of achieving reasonably good quality stereo audio. In this paper, we propose an
improved parametric extraction scheme in the MDCT domain using three different approaches. In
the first scheme we use two reverberated channel instead of the single reverberated channel as
proposed in [14]. This results in better modeling of spatialization cues which is reflected in the
perceptual evaluation. In the second scheme, we use sub-band wise mutually orthogonalized
sum channel and reverberated channels for parameter extraction and synthesis. In the third
scheme, psychoacoustically weighted parameter extraction scheme is introduced. Performance
evaluations of proposed methods are conducted through listening tests.
2. SPATIALIZATION PARAMETER EXTRACTION AND STEREO SYNTHESIS
USING MULTIPLE REVERBERATED CHANNELS
Formation of ’sum’ channel through down-mixing, generating reverberated channels from the
’sum’ channel and parameter extraction are the functions of a parametric stereo encoder. We
follow the methods as used in [14] for in our encoder as described below. The block diagram for
the MDCT domain parametric stereo encoding section is shown in Figure 1. In the first step,
MDCT
Figure 1: Block Diagram of MDCT Domain Parametric Stereo Encoder.

of the stereo channels )(( 1 nx and ))(2 nx are computed. The stereo signals in the MDCT domain
are then down-mixed to form a ‘sum’ channel )(kS . A de-correlated channel )(kSr , is derived
from the sum channel. Additional de-correlated channel is generated by delaying the de-
correlated channel. The frequency domain signals )(),( kSkS r and a delayed version of the de-
correlated channel )(kSrd are used to extract the spatialization parameters.
2.1 Down Mixing to Sum Signal
For the analysis of stereo signal 1024 point MDCT of the input stereo channels are computed. In
order to imitate the features of human auditory system which is more sensitive to lower frequency
bands than higher bands, the MDCT coefficients are grouped into bands with different spectral
resolution. We use a partitioning method same as that suggested by C. Faller in [1] in which the
MDCT coefficients are grouped into 22 non-overlapping frames of different lengths. The partition
details are given in table 1. A sub-band by sub-band energy equalized ‘sum’ signal is obtained
from stereo transforms by down mixing. The MDCT of the
th
j sub-block of the energy equalized
sum signal is given by Equation 1.
)1()},()({)( 21 kallforkXkXckS jj
j
j
+=
)2(
))()((
)))(())(((
2
1
2
21
2
2
2
1
∑
∑
+
+
=
k
jj
k
jj
j
kXkX
kXkX
c
)(1 kX j
and )(2 kX j
are the MDCT coefficients of
th
j sub band. The energy equalization factor
jc given by equation 2 is introduced to make the energy of the ‘sum’ sub-band signal equal to
the average energy of the corresponding stereo sub-bands.
Partition B0 B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11
Boundary 0 8 16 24 32 48 64 80 96 128 160 192
Partition B13 B13 B14 B15 B16 B17 B18 B19 B20 B21 B22
Boundary 224 256 288 320 384 448 512 576 640 768 1024
TABLE 1: Partition boundaries of sub-bands in MDCT domain.
2.2 Synthesis of De-correlated Channels
For parametric coding in MDCT domain, the synthesis parameters are not derived directly from
spatialization parameters like ICLD, ICTD and ICC. Instead, we estimate them using the sum
signal, a de-correlated ‘sum’ signal and a delayed de-correlated signal. The same de-correlated
signals are created in the decoder to synthesize the stereo signal. The de-correlated signal is
computed using the method suggested by J. Breebaart et al., in [11], i.e., by convolving the ‘sum’
signal with an all pass de-correlation filter whose impulse response is given by equation 3.
)3()
)1(22
cos(
2
)(
2
0 N
mm
N
mn
N
nh
N
m
−
+= ∑=
ππ

De-correlated signal is generated by the convolution of ‘sum’ signal and )(nh . The convolution is
implemented in the MDCT domain following the method described in [16]. The MDCT of the de-
correlated ‘sum ’is given by
)4(10)},sin()()cos()({|)12(|)( 12128, −≤≤++= ++ NkkSkSkHkS k
mdct
k
mdctdft
Nu
mdct
r θθ
where )12(8, +kH dft
Nu is the 8N point DFT of up-sampled )(nh . The length of the de-correlation
filter is selected such that L << N for which the error due to aliasing is irrelevant [16]. We use de-
correlation filter of length L=40 for N=1024.
2.3 Parameter Estimation in MDCT Domain
For the parameter extraction, we use ‘sum’ signal )(kS , de-correlated signal )(kSr and delayed
de-correlated signal )(kSrd which is obtained by shifting )(kSr by a certain number of samples.
The MDCT coefficients representing the synthesized stereo channels ))(),(( 21 kXkX jj
are given
by the linear combination
)5(),()()()( kSkSbkSakX j
rd
j
r
j
i
jj
i
j
i ++=
where i=1,2 and 1 ≤ j ≤ 22.
To estimate the synthesis parameters
j
i
j
i ba , and
j
ic the inner product between the transforms
of stereo signals )(),( kSkS r and )(kSrd are computed.
〉〈= )()( kSkXd jj
i
j
i
〉〈= )()( kSkXe j
r
j
i
j
i
)6()()( 〉〈= kSkXf j
rd
j
i
j
i
An approximation of the stereo signal is obtained by adding the projections of )(1 kX j
and
)(2 kX j
on )(),( kSkS j
r
j
and )(kS j
rd as
)7()()()()( kSfkSekSdkP j
rd
j
i
j
r
j
i
jj
i
j
i ++=
Further amplitude scaling of )(kP j
i is done to equalize the energy in the synthesized sub-bands
to that of the original sub-band energy by multiplying with the scaling factors computed as below
)8()()( kPskX j
i
j
i
j
i =
where the scale factors
j
is i=1,2 are given by
)9(
))((
))((
2
2
∑
∑=
k
j
i
k
j
ij
i
kP
kX
s
The synthesis parameters are obtained by multiplying the inner products obtain in the first step
with the corresponding scale factors.
)10(j
i
j
i
j
i
j
i
j
i
j
i
j
i
j
i
j
i fscesbdsa ===
The synthesis parameters are compressed using uniform quantizer and encoded as side
information.

2.4 Stereo Decoding
The block diagram for the decoder is shown in Figure 2. The MDCT coefficients of the
synthesized stereo are reconstructed from the linear combination of the equalized ‘sum’ signal
)(kS , and de-correlated signals )(kSr and )(kSrd . The de-correlated and its delayed signals
are obtained using the same procedure followed in the encoder. Side information forms the
weights for the linear combination.
FIGURE 2: Block diagram of parametric stereo decoder.
2.5 Spatialization Parameter Estimation using Orthogonalized De-correlated Channels
De-correlated channels derived from the ’sum’ channel through filtering are not perfectly
correlated. To further reduce the correlation between them we performed sub-band wise
orthogonalization of the channels in the MDCT domain. The ‘sum’ signal )(kS , de-correlated
signal )(kSr , and delayed de-correlated signal )(kSrd , are modified through Gram-Schmidt
orthogonalization procedure. Orthogonalization of these signals is performed in every sub-band to
make
0)()( =〉〈 kSkS j
r
j
0)()( =〉〈 kSkS j
rd
j
)11(0)()( =〉〈 kSkS j
rd
j
r
where j = 1,2,..,22 denotes the sub-band indices )(),( kSkS j
r
j
and )(kS j
rd the corresponding
orthogonalized ‘sum’, de-correlated and delayed de-correlated signals. We used these signals for
parameter extraction. The decoder is also modified to include the orthogonalization steps
performed in the encoder.
2.6 Stereo Decoding with Orthogonalized De-correlated Channels
The block diagram of the stereo decoder is shown in Figure 3. The decoder receives )(kS j
and
spatialization parameters )(),( kbka j
i
j
i and )(kc j
i . From this, first we obtains the de-correlated
signals )(kS j
r and )(kS j
rd . Then sub-band wise orthogonalization is done and the decoder
synthesizes back the stereo channels as a linear combination of these orthogonalized signals
using the spatialization parameters.

FIGURE 3: Stereo decoder for synthesizing stereo from orthogonalized ‘sum’ and reverberated channels.
2.7 Psychoacoustically Weighted Parameter Estimation Scheme
Masking threshold estimated through psychoacoustic analysis is used for determining
perceptually irrelevant components of audio signals and extensively used in perceptual audio
compression algorithms [17]. A higher masking threshold indicates higher noise masking
capability of the frequency component. Components with lower masking threshold are more
sensitive to quantization noise. We used masking threshold estimate to scale the spatialization
parameters. Low weightage is given for spatialization parameters representing frequency
components with higher masking thresholds since they are less sensitive components.
Coefficients corresponding to lower masking threshold are amplified to give them more
weightage. We used MDCT domain masking threshold estimation scheme presented in [18] for
computing the masking threshold estimate )(kTi of the every audio frame. Scaling factors
)(kmi for frequency components are obtained by uniform linear interpolation of
))](max())(min([ kTkT ii into the range [1.5 0.5] such that minimum weightage is given for
frequency components having maximum masking threshold as given below.
( ) )12(
))(min())(max(
)())(max(
1023)( 





−
−
=
kTkT
kTkT
kN
ii
ii
i
)13(
1023
)(
5.0)(
kN
km i
i +=
This scaling factor )(kmi is used to scale the spatialization parameters in the stereo analysis
stage as given in equation
〉〈= )()()()()( kSkXkmkska jj
ii
j
i
j
i
〉〈= )()()()()( kSkXkmkskb j
r
j
ii
j
i
j
i
)14()()()()()( 〉〈= kSkXkmkskc j
rd
j
ii
j
i
j
i
where j=1,2,..,22 denotes the sub-band indices, i the channel number, )(ks j
i the sub-band wise
energy equalizing factor and )(kmi scaling function obtained from the masking threshold
estimate. The method for stereo synthesis from parameters is same as in the case of stereo
synthesis using multiple reverberated channels.

3. SUBJECTIVE EVALUATION RESULTS
To evaluate the perceptual quality of the encoded audio signal using the proposed algorithm,
listening test was conducted. Six listeners participated in the test. The listeners are asked to
evaluate both the spatial audio quality as well as other audible artifacts. In a MUSHRA test [19],
the listeners had to rate the relative perceptual quality of ten processed items against original
excerpts in a 100-point scale with 5 anchors. Tests are conducted with high quality headphone in
a quiet room. The following items are included in the test.
• The original as the hidden reference.
• A low-pass filtered (cut off frequency of 7 kHz) mono channel derived from the original.
• Stereo audio compressed by MPEG-2 AAC with TNS and M/S stereo enabled at a rate of
64 kbps.
• Stereo signal synthesized using uncompressed ‘sum ’signal and synthesis parameters
estimated with single reverberated channel.
• Stereo signal synthesized using uncompressed ‘sum ’signal and synthesis parameters
estimated with two reverberated channel.
3.1 Perceptual evaluation of parametric stereo generated using multiple reverberated
channels
FIGURE 4: MUSHRA Scores for Average and 95% Confidence Intervals for stereo synthesis.
In the case of stereo synthesized using two reverberated channel, we have considered three
different delay conditions (20 ms, 40 ms and 100 ms) for the second delay channel. Thus, there
are seven clips including the hidden references for each channel in the subjective evaluation. All
clips had a resolution of 16 bits per sample and sampling rate of 44.1 kHz. The list of stereo clips
used for the subjective evaluation is shown in Table 2.
The average MUSHRA scores obtained for the seven version of each clip are shown in Table 3.
The average score for all the clips combined is also given. The synthesized audio with two
reverberated signals performed better than the synthesized signal with single reverberated signal.
The average MUSHRA score for the synthesized signal with reverberated channels having no
delay and 40 ms delay is 93.8 while that of single reverberated channel is 89.3. The performance
of 20 ms delayed reverberated channel and 100 ms delayed channel are also very close to that of
the 40 ms delayed case.

Sl. No. Index Name Origin/Artist
1 abba Abba SQAM Database
2 appla Applause SQAM Database
3 bigye Big Yellow Counting Crows
4 charl Charlies Danny O’Keefe
5 cymba Cymbal Radiohead
6 eddie Eddie Rabbit SQAM Database
7 flori Florida Sequence Pre-echo test case
8 helic Helicopter bAdDuDeX
9 orche Orchestra Dave Mathews band
10 ordin Ordinary World Duran Duran
TABLE 2: List of excerpts used in the subjective listening test.
Name Original Low
pass
Mono
MPEG 2
AAC
Synth
Uncomp-
ressed
Synth.
Delay 20
Synth.
Delay 40
Synth.
Delay 40
Abba 100 47 66 86.5 91.7 93.2 95.5
Applause 100 45.5 67.2 87.8 93 94.3 93.3
Big Yellow 100 48 65.5 92 91.3 96 94.3
Charlies 100 46.2 66.8 86.5 93 93.8 91
Cymbal 100 46.7 63.5 87.5 93.2 92.5 91.5
Eddie Rabbit 99.3 50.5 63.7 90 96.3 94.5 92.7
Florida Sequence 100 44.2 64.3 88.2 92.2 93 92.8
Helicopter 99.2 46.8 67.8 92 96.5 91.5 95.2
Orchestra 100 43.3 64.2 89.5 90 93 93.8
Ordinary World 99.7 45.3 64.8 92.8 92.5 96.7 94.2
All Items 99.8 46.4 65.4 89.3 93 93.8 93.4
TABLE 3: Average MUSHRA scores for test clips generated using multiple reverberated channels.
3.2 Perceptual evaluation of parametric stereo generated using orthogonalized signals
To evaluate the performance of stereo synthesized using orthogonalized de-correlated channels
subjective evaluation was conducted with following audio clips.
• A low-pass filtered (cut off frequency of 7 kHz) mono channel derived from the original.
• Stereo audio compressed by MPEG-2 AAC at a rate of 32 kbps.
• Stereo signal synthesized using energy equalized ‘sum’ signal, de-correlated signal and
its delayed version (with delay of 40 samples) weighted by synthesis parameters.
• Stereo signal synthesized using orthogonalized and energy equalized ‘sum’ signal, de-
correlated signal and its delayed version (with delay of 40 samples) weighted by
synthesis parameters.

FIGURE 5: MUSHRA Scores for Average and 95% Confidence Intervals for audio clips.
Name Original Low pass
mono
MPEG 2 AAC Reverb. delay Orthogonal
Delay
Abba 100 38.9 59.8 83.5 89.8
Applause 100 29.9 52.2 80.5 87.1
Big Yellow 98.8 28.8 50.2 86 90
Charlies 99.6 39 57.5 88.1 89.5
Cymbal 99 36.9 52.6 82.5 89.4
Eddie Rabbit 99.4 30 45.9 84.1 90.8
Florida
Sequence
100 33.8 45.2 81.8 86.6
Helicopter 99.8 31.4 56.1 86 89.9
Orchestra 100 33.6 43 82.9 87.1
Ordinary
World
97.8 32.2 47.4 83.2 89
All Items 99.4 33.4 51 83.9 89.9
TABLE 4: Average MUSHRA scores of audio clips synthesized using orthogonalized signals.
Stereo signal synthesized using energy equalized ‘sum’ signal, de-correlated signal and its
delayed version(with delay of 40 samples) weighted by synthesis parameters. Stereo signal
synthesized using orthogonalized and energy equalized ‘sum’ signal, de-correlated signal and its
delayed version(with delay of 40 samples) weighted by synthesis parameters. The average
MUSHRA scores assigned by the listeners for these clips are shown in Table 4. The average
score for all the clips combined is also given. MUSHRA scores for average and 95% confidence
intervals are plotted in Figure 5.
Results clearly show that orthogonalizing the ‘sum’ and reverberated channels for spatialization
parameter extraction and stereo synthesis have resulted in improving the perceptual quality of the
synthesized audio. The average MUSHRA score for the synthesized stereo has increased from

83.9 to 89.9 when orthogonalized signals were used for parameter extraction and stereo
synthesis.
3.3 Perceptual evaluation of parametric stereo generated using scaled parameters
Test audio clips were generated by scaling spatialization parameters extracted using estimated
masking threshold values. This perceptual weighting of parameters is done for the stereo
synthesis using multiple reverberated channels with and without orthogonalization. We used the
following items for listening test:
• Stereo audio compressed by MPEG-2 AAC at a rate of 32 kbps.
synthesis parameters.
its delayed version (with delay of 40 samples) with scaled parameters.
• Stereo signal synthesized using orthogonalized and energy equalized ‘sum’ signal, de-
correlated signal and its delayed version (with delay of 40 samples) with scaled
parameters.
The average MUSHRA scores obtained for different test audio clips and average score obtained
for all clips are shown in Table 5. MUSHRA scores for average and 95% confidence intervals are
FIGURE 6: MUSHRA Scores for average and 95% confidence Intervals obtained for synthesized audio clips
using scaled spatialization parameters.
plotted in Figure 6 The results of the listening test shows that the perceptual quality of the
synthesized stereo using two reverberated channels was 86.7 whereas including masking
threshold also to scale the spatialization parameters resulted in an average score of 87.6.
3.4 Discussion
It is not surprising that the quality of the encoder increases as the number of parameters is
increased. Use of an additional delayed reverberated channel for stereo analysis and synthesis
increases the perceptual quality at the cost of increased side information rate. The delayed signal
helps the parametric model to capture the late reverberated part of the original signal. The
reverberation filter length is 40 samples which is approximately equal to 1 ms and a delay of 40
samples effectively extends the filter length to 80 samples and that results in better modeling of
the spatial cues. The number of parameters representing spatialization cues is increased to three
per sub-band for each channel and this result in a better approximation for the stereo audio.

Name Original Low pass
mono
MPEG 2
AAC
Synth.
Uncompres
sed
Synth.
Delay 40
Perceptually
Weighted
Abba 99.1 32.8 60.6 84 84.9 89.4
Applause 96.8 33.4 49.5 86.8 83.9 82.4
Big Yellow 96.4 32.9 54.2 83.6 86.6 87.1
Charlies 98 33.4 53.4 82.6 86.9 86
Cymbal 100 33.1 52.2 84.1 82.9 87.1
Eddie
Rabbit
98.9 32.1 46.8 85.2 89.4 91.1
Florida
Sequence
97.1 31.5 46.8 84.6 85 85.4
Helicopter 94.1 33.2 51 87.2 91.4 94.8
Orchestra 97.8 31.6 46.1 84 88 86.8
Ordinary
World
97.2 32.5 52.6 86.2 87.6 85.9
All Items 97.5 32.6 51.3 84.8 86.7 87.6
TABLE 5: Average MUSHRA scores of synthesized audio clips using scaled Spatialization parameters.
Perceptual tests shows better quality for stereo synthesized using additional reverberated
channel delayed by 40 samples with an average MUSHRA score of 93.4 whereas that for stereo
synthesized using single reverberated channel was 89.3. When we orthogonalize the de-
correlated signals we expect better spatial modeling at the cost of additional computation.
MUSHRA test reveals a better performance of stereo synthesized using orthogonalized signals
with an average score of 89.9 against the average score of 83.3 obtained for stereo synthesized
non orthogonalized reverberated channels. Stereo synthesis using perceptually scaled
spatialization parameters produce marginal improvement in perceived quality at the cost of
increased complexity. When evaluated simultaneously, the average MUSHRA score is 87.6 for
scaled parameter synthesis while that of multiple de-correlated channels is 86.7.
4. CONCLUSION
Methods for parametric stereo coding in MDCT domain using sum and two de-correlated
channels have been introduced. Three methods are used for estimating spatialization
parameters. It can be seen that stereo synthesized using ‘sum’ and two de-correlated channels
has a better perceptual quality than that synthesized using ‘sum’ and a single de-correlated
channel. The quality of the encoder has increased when orthogonalized signals were used for
parameter extraction and stereo synthesis. Orthogonalization makes the three signals used in
parameter extraction and stereo synthesis independent of each other. But the computational
complexity of encoder as well as decoder will be increased due to the additional orthogonalization
process. In method 3, spatialization parameters were further modified using scaling functions
obtained from masking threshold and results in marginal improvement in the perceptual quality of
the synthesized audio.
5. REFERENCES
[1] C. Faller, “Parametric Coding of Spatial Audio,” Swiss Federal Institute of Technology
Lausanne (EPFL), PhD Thesis, No. 3062, 2004.

[2] D. Yang, H. Ai, C. Kyriakakis, ans C.C. J. Kuo, “An inter channel redundancy removal
approach for high quality multichannel audio compression,” in AES convention, Los
Angeles, CA, Sept 2000.
[3] S. Kuo and J.D. Johnston, “A Study of Why Cross Channel Prediction is Not Applicable to
Perceptual Audio Coding," IEEE Sig. Proc. Letters, vol. 8, No. 9, pp 245-247, Sep. 2001.
[4] J. Herre, et.al, “The reference Model Architecture for MPEG Spatial Audio Coding," in 118th
AES convention, Barcelona, Spain May 2005, Preprint 6447.
[5] J.D. Johnston, and A.J. Ferreira, “Sum Difference Stereo Transform Coding,” in Proc. IEEE
ICASSP-92, San Francisco, vol. 2, pp. 569-572, March 1992.
[6] Christian R. Helmrich, Pontus Carlsson, Sascha Disch, Bernd Edler, Johannes Hilpert,
Matthias Neusinger, Heiko Purnhagen, Nikolaus Rettelbach, Julien Robilliard, and Lars
Villemoes, “Efficient Transform Coding Of Two-Channel Audio Signals By Means Of
Complex-Valued Stereo Prediction,” in Proc. IEEE ICASSP-2011, pp. 497-500, 2011.
[7] Christof Faller, and Frank Baumgarte, “Binaural Cue Coding: A Novel and Efficient
Representation of Spatial Audio,” in Proc. IEEE ICASSP-2002, vol: 2, pp. II-1841 - II-1844,
2002.
[8] F. Baumgarte, and C. Faller,“Binaural Cue Coding-part I : Psychoacoustic fundamentals and
Design Principles,” in IEEE Trans. on Speech and Audio Proc., vol. 11, No. 6, pp. 509-519,
June 2003.
[9] F. Baumgarte, and C. Faller,“Binaural cue coding-part II : Schemes and applications,” in
IEEE Trans. on Speech and Audio Proc., vol. 11, No. 6, pp. 520-531, June 2003.
[10] C. Faller, “Parametric Multichannel Audio Coding: Synthesis of Coherence Cues," IEEE
Trans. Speech and Audio Proc., vol. 14, No. 1, pp. 1-12, Jan. 2006.
[11] J. Breebaart, et al.,“Parametric Coding of Stereo Audio,” in EURASIP Journal on Applied
Signal Processing, vol 2005, No. 9, pp 1305 - 1322, June 2005.
[12] A. Kohlrausch, “Auditory filter shape derived from binaural masking experiments," J. Acous.
Soc. America, vol. 84, no. 2, pp. 573-583, 1988. 16
[13] B. R. Glasberg and B.C.J. Moore, “Derivation of auditory filter shapes from notched-noise
data," Hearing Research, vol. 47, no. 1-2, pp . 103-138, 1990.
[14] K. Suresh, and T. V. Sreenivas, “MDCT Domain Analysis and Synthesis of Reverberation for
Parametric Stereo Audio,” in AES 123th Convention, 2007 October 5-8, New York.
[15] K. Suresh, and T. V. Sreenivas, “Parametric stereo coder with only MDCT domain
computations,” IEEE International Symposium on Signal Processing and Information
Technology, pp. 61-64, December 2009.
[16] K. Suresh and T. V. Sreenivas, “Linear Filtering in DCT-IV/DST-IV and MDCT/MDST
Domain”, Signal Processing, vol 89, Issue 6, pp 1081-1089, June 2009.
[17] T. Painter, and A. Spanias, “Perceptual Coding of Digital Audio", Proc. IEEE, vol. 88, no 4,
pp. 451-513, 2000.
[18] K Suresh and T. V. Sreenivas, “Direct MDCT Domain Psychoacoustic Modeling”, IEEE
International Symposium on Signal Processing and Information Technology, pp. 742-747,
December 2007.

[19] ITU/ITU-R BS 1534. Method for subjective assessment of intermediate quality level of
coding systems, 2001.

Spatialization Parameter Estimation in MDCT Domain for Stereo Audio

More Related Content

What's hot (18)

Similar to Spatialization Parameter Estimation in MDCT Domain for Stereo Audio (20)

Recently uploaded (20)

Spatialization Parameter Estimation in MDCT Domain for Stereo Audio