In large MP3 databases, files are typically generated with different parameter settings, i.e., bit rate and sampling rates. This is of concern for MIR applications, as encoding difference can potentially confound meta-data estimation and similarity evaluat
MelFrequencyCepstralCoef cients:AnEvaluationofRobustnessofMP3
EncodedMusic
SigurdurSigurdsson,KaareBrandtPetersenandTueLehn-Schiøler
InformaticsandMathematicalModelling
TechnicalUniversityofDenmarkRichardPetersensPlads-Building321DK-2800Kgs.Lyngby-Denmark{siggi,kbp}@imm.dtu.dk
InlargeMP3databases,Abstract
lesaretypicallygeneratedwithdifferentparametersettings,i.e.,bitrateandsamplingrates.ThisisofconcernforMIRapplications,asencodingdif-ferencecanpotentiallyconfoundmeta-dataestimationandsimilarityevaluation.Inthispaperwewilldiscussthein- uenceofMP3codingfortheMelfrequencycepstralcoe- cients(MFCCs).ThemainresultisthatthewidelyusedsubsetoftheMFCCsisrobustatbitratesequalorhigherthan128kbits/s,fortheimplementationswehaveinvesti-gated.However,forlowerbitrates,e.g.,64kbits/s,theim-plementationoftheMel lterbankbecomesanissue.Keywords:Melfrequencycepstralcoef cients,MFCC,ro-bustness,MP3.
1.Introduction
TheuseofMelfrequencycepstralcoef cients(MFCCs)for
musicinformationretrievalhasbecomestandardsincetheseminalpaper[4]in1997.ButonlylittleefforthasbeenputintoinvestigatingtheapplicabilityoftheMFCC’sasfea-turesformusic,with[6]asarareexception.InthispaperweinvestigatehowMP3encodingofmusic lesisin uenc-ingthesignalinformationcontentoftheMFCC’s.
2.MelFrequencyCepstralCoef cients
WewillusetheIntelligentsoundimplementation(ISP)to
explainthecomputationofMFCCs.Firstthemusicsignalisdividedintoshorttimewindows,wherewecomputethediscreteFouriertransform(DFT)ofeachtimewindowforthediscrete-timesignalx(n)withlengthN,givenby
NX(k)=
1w(n)x(n)exp( j2πkn/N)(1)
n=0
fork=0,1,...,N 1,wherekcorrespondstothefre-quencyf(k)=kfs/N,fsisthesamplingfrequencyin
Permissiontomakedigitalorhardcopiesofallorpartofthisworkfor
personalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforpro torcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthe rstpage.
c2006UniversityofVictoriaHertzandw(n)isatime-window.Here,wechosethepopu-larHammingwindowasatimewindow,givenbyw(n)=0.54 0.46cos(πn/N),duetocomputationalsim-plicity.
Themagnitudespectrum|X(k)|isnowscaledinbothfrequencyandmagnitude.First,thefrequencyisscaledlog-arithmicallyusingtheso-calledMel lterbankH(k,m)andthenthelogarithmistaken,X′(m)=ln
giving
N
1
|X(k)|·H(k,m) (2)k=0
forandmlarM= 1,N2,....,M,whereMisthenumberof lterbanksas
ltersde nedThebyMelthe ltercenterbankfrequenciesisacollectionfoftriangu-c(m),written0
forf(k)<fc(m 1)
f(k) fc(m 1)H(k,m)=
cforfc(m 1)≤f(k)<fc(m)f(k) fcc(m+1)ccforfc(m)≤f(k)<fc(m+1)0
forf(k)≥fc(m+1).
(3)
Thecenterfrequenciesofthe lterbankarecomputedbyapproximatingtheMelscalewith
φ=2595log10(f
+1),
(4)
whichisacommonapproximation.Notethatthisequation
isnon-linearforallfrequencies.Thena xedfrequencyres-olutionintheMelscaleiscomputed,correspondingtoalog-arithmicscalingoftherepetitionfrequency,using φ=(φmax φmin)/(M+1)whereφmaxisthehighestfrequencyofthe lterbankontheMelscale,computedfromfmaxusingequation(4),φministhelowestfrequencyinMelscale,havingacorrespondingfmin,andMisthenum-berof lterbanks.ThevaluesfortheISPimplementationisfmax=11.025kHz,fmin=0Hz,andM=30.ThecenterfrequenciesontheMelscalearegivenbyφc(m)=m· φform=1,2,...,M.ToobtainthecenterfrequenciesinHertz,weapplytheinverseofequation(4),givenbyfc(m)=700(10φc(m)/2595 1),whichareinsertedintoequation(3)togivetheMel lterbank.Finally,theMFCCsareobtainedbycomputingtheDCTofX′(m)using
c(l)=
MX′(m)cos(l(m ))(5)
m=1
An evaluation of visua... 暂无评价 8页 免费 Noise...MFCC (Mel Frequency Cepstral Coefficients) o One ...robustness o MVDR o MHEC o Multitaper o Improve...
Speech reconstruction from mel frequency cepstral coefficients and pitch This paper presents a novel low complexity, frequency domain algorithm for reconstruction...
The well-known mel-frequency cepstral coefficients (MFCCs) [1], though ...Although some of the loss caused by the imprecision and low robustness of ...
evaluation corpus, which contains audio-visual ...(GMM) of mel frequency cepstral coefficients (...order to improve robustness against channel mismatch...
an additional feature to cepstral coefficients in ...evaluation corpus, a conversational cellular phone ...lters – SSC(3) : mel frequency scale, ...
robustness in low signal-to-noise ratio setups. ...among which Mel frequency cepstral coefficients [5...evaluation of the bird species classification step:...
an additional feature to cepstral coefficients in ...evaluation corpus, a conversational cellular phone ...lters – SSC(3) : mel frequency scale, ...
热门文档