Date post: | 24-May-2015 |
Category: |
Technology |
Upload: | a3labdsp |
View: | 286 times |
Download: | 0 times |
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Hybrid Reverberator Using Multiple ImpulseResponses for Audio Rendering Improvement
Andrea Primavera1, Stefania Cecchi1, Francesco Piazza1, Junfeng Li2, and
Yonghong Yan2
1 A3lab - DII - Universita Politecnica delle Marche -Ancona - ITALY
2 Institute of Acoustics, Chinese Academy ofSciences - Beijing - CHINA
IIHMSP, October 2013, Beijing, China.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 1/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
1 Audio Rendering Based on Multiple Impulse ResponsesIntroductionState of the artProposed algorithm
2 Proposed algorithmAnalysis of reverberation tailSynthesis of the reverberation effectReal-time Reproduction of Moving Listener Position
3 Experimental ResultsExperimental SetupReverberation TimeClarity IndexSubjective Analysis
4 ConclusionConclusionQuestionsBibliography
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 2/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
IntroductionState of the artProposed algorithm
Linear convolution is a widely used operation typically employed for audiorendering purpose aiming to reproduce the reverberation effect generatedwhen a sound is produced within an enclosed space.
LINEARCONVOLUTION
STATICPROCEDURE
It allows to reproduce only the acoustic effect produced taking intoaccount a specific sound source with the relative receiver position.
One of the main problems
• Time varying convolution to simulate the moving receiverpositions performing IRs interpolation [1].
• A large impulse response (IR) database is needed.
Solution
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 3/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
IntroductionState of the artProposed algorithm
Linear convolution is a widely used operation typically employed for audiorendering purpose aiming to reproduce the reverberation effect generatedwhen a sound is produced within an enclosed space.
LINEARCONVOLUTION
STATICPROCEDURE
It allows to reproduce only the acoustic effect produced taking intoaccount a specific sound source with the relative receiver position.
One of the main problems
• Time varying convolution to simulate the moving receiverpositions performing IRs interpolation [1].
• A large impulse response (IR) database is needed.
Solution
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 3/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
IntroductionState of the artProposed algorithm
Linear convolution is a widely used operation typically employed for audiorendering purpose aiming to reproduce the reverberation effect generatedwhen a sound is produced within an enclosed space.
LINEARCONVOLUTION
STATICPROCEDURE
It allows to reproduce only the acoustic effect produced taking intoaccount a specific sound source with the relative receiver position.
One of the main problems
• Time varying convolution to simulate the moving receiverpositions performing IRs interpolation [1].
• A large impulse response (IR) database is needed.
Solution
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 3/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
IntroductionState of the artProposed algorithm
PROBLEM: Large impulse responses database required high memoryusage.
In [2] a database reduction procedure for auralization purpose with movinglistener position has been proposed:
• Early reflections contain most of the information regarding thelocation of the sound source and receiver.
• Late reflections gives more information about room properties(e.g., size, geometry, materials) [3].
• The information in the late reverberation tail is largelyredundant across multiple impulse responses recorded in thesame space.
Consideration
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 4/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
IntroductionState of the artProposed algorithm
PROBLEM: Large impulse responses database required high memoryusage.
In [2] a database reduction procedure for auralization purpose with movinglistener position has been proposed:
• Mixing time evaluation to discriminate late from earlyreflections.
• Approximation of the reverberation tail of the whole IRdatabase as stochastic process (i.e., white noise).
Metodology
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 5/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
IntroductionState of the artProposed algorithm
PROBLEM: Large impulse responses database required high memoryusage.
Taking into account the procedure described in [2] a novel metho-dology has been proposed considering the advantages introduced byhybrid reveberation structure [4] [5].
• It is possible to approximate the convolution operation usingrecursive structure (i.e., IIR filters).
• This procedure allows to further reduce the database dimensionwith respect to [2].
• The employed structures permits to decrease the computationalload required to perform the real-time auralization.
Proposed Solution
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 6/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Analysis of reverberation tailSynthesis of the reverberation effectReal-time Reproduction of Moving Listener Position
Three are the main phases of the approach presented for the reproductionof moving listener position exploiting a hybrid reverberator structure:
1 Analysis of reverberation tail:Generate a prototype representing the database average reveberationtail.
2 Synthesis of the reverberation effect:Approximate the reverberation tail prototype exploiting an hybridreverberation algorithm [4] [5].
3 Real-time reproduction of moving listener position:Reverberation effect reproduction using the hybrid reverberationstructure (mixed FIR/IIR filter network).
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 7/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Analysis of reverberation tailSynthesis of the reverberation effectReal-time Reproduction of Moving Listener Position
1 Analysis of reverberation tail:
• Mixing time analysis:The partitioning of early from late reflections has been performedexploiting gaussianity [4] [6] and phase evolution estimators [7].
• Prototype evaluation:The reverberation tail prototype is computed as a mean of the IRsdatabase after the maximum mixing time tm.
htail =1
N
Lm∑t=tm
hn(t) (1) hn: database IRs
• Scaling operation:In order to simulate the distance among the source and the differentlistereners position a scaling factor is evaluated.
Sm =RMS(hs )
RMS(hm)=
√1
Lm
Lm∑t=tm
h2s (t)√
1Lm
Lm∑t=tm
h2m(t)
(2)
hs : synthesized IRhm: original IR
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 8/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Analysis of reverberation tailSynthesis of the reverberation effectReal-time Reproduction of Moving Listener Position
2 Synthesis of the reverberation effect:
+EARLY
REFLECTIONS
DEVICE
LATE
REFLECTIONS
DEVICE
DELAY ×x[n] y[n]
gain
Hybrid reverberator block diagram for the single channel case.
Based on the convolution with areal IR for the reproduction of theearly echoes.
Early reflections device
Based on IIR filters network (e.g.,comb and/or all-pass) and a FDNmatrix [8] for the simulation of thereverberation tail.
Late reflections device
LBCF
+x[n]
LBCF
LBCF
LBCF
AP AP
A
++
++
++
+
+y[n]
NAP filters
NLBCF filters
Late reflections device block diagram
for the single channel case.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 9/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Analysis of reverberation tailSynthesis of the reverberation effectReal-time Reproduction of Moving Listener Position
2 Synthesis of the reverberation effect:
+EARLY
REFLECTIONS
DEVICE
LATE
REFLECTIONS
DEVICE
DELAY ×x[n] y[n]
gain
Hybrid reverberator block diagram for the single channel case.
An automatic procedure allows toset the parameters of hybrid re-verberator in order to emulate areal environment starting from itsimpulse.
FIR TO IIR APPROXIMATION
Autotuning procedure
LBCF
+x[n]
LBCF
LBCF
LBCF
AP AP
A
++
++
++
+
+y[n]
NAP filters
NLBCF filters
Late reflections device block diagram
for the single channel case.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 10/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Analysis of reverberation tailSynthesis of the reverberation effectReal-time Reproduction of Moving Listener Position
2 Synthesis of the reverberation effect:
Two are the main phases of the autotuning procedure:
Evaluation of the mixing time to set theearly reflection device:
• Gaussianity estimators:Similarities between IR behaviorand gaussian noise can be foundin late reflections.Kurtosis and MAD/SD ratiohave been used.
• Phase distortion evaluation:The unwrapped phase of the IRtends to become not linear withlate reflections evolution.
Early Reflections Partitioning An offline adaptation procedure, ba-sed on SPSA [9], has been used toiteratively find the IIR parameters.A single loss function computed in cep-stral domain [10] has been adopted inthe minimization procedure.
L = max
max
√√√√√ K∑
i=1
M∑j=1
[Tr (i, j) − Ta(i, j)]2
where:• Tr is a matrix representing the
MFCC derived from the real IR.
• Ta is the MFCC obtained by theartificial IR.
Late Reflections Analysis
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 11/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Analysis of reverberation tailSynthesis of the reverberation effectReal-time Reproduction of Moving Listener Position
3 Real-time reproduction of moving listener position:
The movement reproduction is generate changing the reverberator para-meters as a function of the listener position:
• Early reflection device: The filter coefficientsare obtained as alinear or bilinear interpolation of the impulse response, with relationto the number of the involved IR.
• Late reflection device: The coefficients are fixed reproducing thesame reverberation tail for all the different positions.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 12/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Experimental SetupReverberation TimeClarity IndexSubjective Analysis
The effectiveness of the presented technique has been proved taking intoaccount account the IR database of a real environment (i.e., St. MargaretsChurch in York [11]).
As reported in [11], a total of 18IRs has been derived using:
• A logarithmic sweep signalexcitation (20 Hz - 22 kHz).
• Sample rate of 96kHz.
• A Soundfield SPS422Bmicrophone.
• A Genelec S30D loudspeaker.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 13/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Experimental SetupReverberation TimeClarity IndexSubjective Analysis
The reverberation time as a function of frequency has been analyzed inorder to provide an objective evaluation between real and synthesize IRs.
Energy Decay Relief (EDR) of one of
the eighteen real IR.
Energy Decay Relief (EDR) of
artificial IR.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 14/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Experimental SetupReverberation TimeClarity IndexSubjective Analysis
The reverberation time as a function of frequency has been analyzed inorder to provide an objective evaluation between real and synthesize IRs.
Mean difference in reverberation time between the measured and synthesized IRs
exploiting (a) the proposed approach and (b) the technique described in [2].
Since the obtained errors are comparable, the effectiveness of the proposedtechnique in time frequency behaviors reproduction is confirmed.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 15/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Experimental SetupReverberation TimeClarity IndexSubjective Analysis
Another parameter employed in objective analysis is clarity (C50 and C80):
C50
Mean real Mean synth Mean err Std errProposed approach 1.75 2.28 0.52 1.09
Approach presented in [2] 1.75 0.10 1.65 1.97
C80Proposed approach 4.23 5.49 1.26 1.06
Approach presented in [2] 4.23 3.41 0.81 1.29
Clarity measures: mean and the standard deviation (STD) error computed asdifference in corresponding receiver positions of the synthesized and measured IRs
exploiting the proposed approach and the one described in [2].
The similar values obtained as mean and standard deviation evaluationconfirms the validity of the approach.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 16/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
Experimental SetupReverberation TimeClarity IndexSubjective Analysis
Informal listening tests have been performed:
• The movement of the listener position along one dimension has beensimulated.
• The effectiveness of the approach has been confirmed since listenersare not able to hear any difference between the presented approachand the one described in [2].
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 17/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
ConclusionQuestionsBibliography
In conclusion:
• A novel approach for the reproduction of moving listener positionexploiting time variant hybrid reverberation algorithm has beenpresented.
• As confirmed in several papers the employment of IIR filter networkfor the approximation of convolution operation allows to reduce thecomputational cost required in the auralization operation, moreoverthe approach also allow to decrease the IR database size reducingthe information required for the late reflection reprodution.
• The effectiveness of the approach has been proved taking intoaccount a real IR database providing comparison with the existingstate-of-art techniques in terms of objective and subjective measures.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 18/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
ConclusionQuestionsBibliography
QUESTIONS?
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 19/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
ConclusionQuestionsBibliography
B. Dalenback and M. Stromberg, “Real time walk-throughauralization - the first year,” in Proc. Institute of Acoustics,Amsterdam, NL, Mar. 2006.
R. Stewart and M. Sandler, “Generating a spatial averagereverberation tail across multiple impulse responses,” in Proc. 35thAudio Engineering Society Conference, London, UK, Dec. 2009.
B. Blesser, “An interdisciplinary synthesis of reverberationviewpoints,” J. Audio Eng. Soc., vol. 49, no. 10, pp. 867–903, Oct.2001.
A. Primavera, S. Cecchi, P. Peretti, L. Romoli, and F. Piazza, “AnAdvanced Implementation of a Digital Artificial Reverberator,” inProc. 130th Audio Engineering Society Convention, London,UK,May 2011.
A. Primavera, M. Gasparini, S. Cecchi, L. Romoli, and F. Piazza,“Hybrid Reverberation Algorithm: a Practical Approach,” inAIA-DAGA Conference, Merano, Italy, Mar. 2013.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 20/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
ConclusionQuestionsBibliography
R. Stewart and M. Sandler, “Statisical measures of early reections ofroom impulse responses,” in in DAFX 07), Bordeaux, France, Sep.2007.
G. Defrance and J. Polack, “Measuring the mixing time inauditoria,” in Proc. 155th Meeting of the Acoustical Society ofAmerica), vol. 49, Jun 2001, pp. 867–903.
J. Jot, “Digital Delay Networks for designing artificial reverberators,”in Proc. 90th Audio Engineering Society Convention, Paris, Feb1991.
J. Spall, “Implementation of the Simultaneous PerturbationAlgorithm for Stochastic Optimization,” in IEEE Transactions onAerospace and Electronic Systems, vol. 34, 1998, pp. 817–823.
S. Heise, M. Hlatky, and J. Loviscach, “Automatic Adjustment ofOff-the-Shelf Reverberation Effects,” in Proc. 126th AudioEngineering Society Convention, Munich, Germany, May 2009.
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 21/22
Audio Rendering Based on Multiple Impulse ResponsesProposed algorithm
Experimental ResultsConclusion
ConclusionQuestionsBibliography
“OpenAIR, Audiolab, University of York.” [Online]. Available:http://www.openairlib.net/
Andrea Primavera Hybrid Reverberator Using Multiple Impulse Responses for Audio Rendering Improvement 22/22