Proceedings Fonetik 2009 - Institutionen för lingvistik

More documents

Recommendations

Info

Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm UniversityOn extending VTLN to phoneme-specific warping inautomatic speech recognitionDaniel Elenius and Mats BlombergDepartmen of Speech, Music and Hearing, KTH, StockholmAbstractPhoneme- and formant-specific warping hasbeen shown to decrease formant and cepstralmismatch. These findings have not yet been fullyimplemented in speech recognition. This paperdiscusses a few reasons how this can be. A smallexperimental study is also included where phoneme-independentwarping is extended towardsphoneme-specific warping. The results of thisinvestigation did not show a significant decreasein error rate during recognition. This isalso in line with earlier experiments of methodsdiscussed in the paper.IntroductionIn ASR, mismatch between training and testconditions degrades the performance. Thereforemuch effort has been invested into reducing thismismatch using normalization of the inputspeech and adaptation of the acoustic modelstowards the current test condition.Phoneme-specific frequency scaling of aspeech spectrum between speaker groups hasbeen shown to reduce formant- (Fant, 1975) andcepstral- distance (Potamianos and Narayanan,2003). Frequency scaling has also been performedas a part of vocal tract length normalization(VTLN) to reduce spectral mismatchcaused by speakers having different vocal tractlengths (Lee and Rose 1996). However, in contrastto findings above this scaling is normallymade without regard to sound-class. How comethat phoneme-specific frequency scaling inVTLN has not yet been fully implemented inASR (automatic speech recognition) systems?Formant frequency mismatch was reducedby about one-half when formant- andvowel-category- specific warping was appliedcompared to uniform scaling (Fant, 1975). Alsophoneme-specific warping without formant-specificscaling has been beneficial interms of reducing cepstral distance (Potamianosand Narayanan, 2003). In the study it was alsofound that warp factors differed more betweenphonemes for younger children than for olderones. They did not implement automatic selectionof warp factors to be used during recognition.One reason presented was that the gain inpractice could be limited by the need of correctlyestimating a large number of warp factors.Phone clustering was suggested as a method tolimit the number of warping factors needed toestimate.One method used in ASR is VTLN, whichperforms frequency warping during analysis ofan utterance to reduce spectral mismatch causedby speakers having different vocal tract lengths(Lee and Rose, 1996). They steered the degree ofwarping by a time-independent warping-factorwhich optimized the likelihood of the utterancegiven an acoustic model using the maximumlikelihood criterion. The method has also beenfrequently used in recognition experiments bothwith adults and children (Welling, Kanthak andNey, 1999; Narayanan and Potamianos, 2002;Elenius and Blomberg, 2005; Giuliani, Gerosaand Brugnara 2006). A limitation with this approachis that time-invariant warping results inall phonemes as well as non-speech segmentssharing a common warping factor.In recent years increased interest has beendirected towards time-varying VTLN (Miguelet.al., 2005; Maragakis et.al., 2008). The formermethod estimates a frame-specific warpingfactor during a memory-less Viterbi decodingprocess, while the latter method uses a two-passstrategy where warping factors are estimatedbased on an initial grouping of speech frames.The former method focuses on revising the hypothesisof what was said during warp estimationwhile the latter focuses on sharing the samewarp factor within each given group. Phoneme-specificwarping can be implemented tosome degree with either of these methods. Eitherby explicitly forming phoneme-specific groupsor implicitly by estimating frame-specific warpfactors.However, none of the methods above presentsa complete solution for phoneme-specificwarping. One reason is that more then one instantiationof a phoneme can occur far apart intime. This introduces a long distance depend-144
Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm Universityency due to a shared warping factor. For theframe-based method using a memory-lessViterbi process this is not naturally accountedfor.A second reason is that in an unsupervisedtwo-pass strategy initial mismatch causes recognitionerrors which limit the performance.Ultimately initial errors in assigning frames togroup-identities will bias the final recognitionphase towards the erroneous identities assignedin the first pass.The objective of this paper is to assess theimpact of phoneme-specific warping on anASR-system. First a discussion is held regardingissues with phoneme-specific warping. Then anexperiment is set up to measure the accuracy of asystem performing phoneme-specific VTLN.The results are then presented on a connected-digittask where the recognizer wastrained for adults and evaluated on children’sspeech.Phoneme-specific VTLNThis section describes some of the challenges inphoneme-specific vocal tract length normalization.Selection of frequency warping functionIn (Fant, 1975) a case was made forvowel-category and formant- specific scaling incontrast to uniform scaling. This requires formanttracking and subsequent calculations offormant-specific scaling factors, which is possibleduring manual analysis. Following anidentical approach under unsupervised ASRwould include automatic formant tracking,which is a non-trivial problem without a finalsolution (Vargas at.al, 2008).Lee and Rose (1996) avoided explicitwarping of formants by performing a commonfrequency warping function for all formants.Since the function is equal for all formants, noformant-frequency estimation is needed whenapplying this method. The warping function canbe linear, piece-wise linear or non-linear. Uniformfrequency scaling of the frequency intervalof formants is possible using a linear orpiece-wise linear function. This could also beextended to a rough formant scaling, using anon-linear function, under the simplified assumptionthat the formant regions do not overlap.This paper is focused on uniform scaling ofall formants. For this aim a piece-wise linearwarping function is used, where the amount ofwarping is steered by a warping-factor.Warp factor estimationGiven a specific form of the frequency warpingto be performed, a question still remains of thedegree of warping. In Lee and Rose (1996) thiswas steered by a common warping factor for allsound-classes. The amount of warping was determinedby selecting the warping-factor thatmaximized the likelihood of the warped utterancegiven an acoustic model. In the generalcase this maximization lacks a simple closedform and therefore the search involves an exhaustivesearch on a set of warping factors.An alternative to warp the utterance is toperform a transform of the model parameters ofthe acoustic model towards the utterance.Thereby a warp-specific model is generated. Inthis case, warp factor selection amounts to selectingthe model that best fit data, which is astandard classification problem. So given a setof warp-specific models one can select themodel that results in the maximum likelihood ofthe utterance.Phoneme-specific warp estimationLet us consider extending the method above to aphoneme-specific case. Instead of a scalarwarping factor a vector of warping factors can beestimated with one factor per phoneme. The taskis now to find the parameter vector that maximizesthe likelihood of the utterance given thewarped models. In theory this results in an exhaustivesearch of all combinations of warpingfactors. For 20 phonemes with 10 warp candidates,this amounts to 10 20 likelihood calculations.This is not practically feasible and therebyan approximate method is needed.In (Miguel et.al, 2005) a two-pass strategywas used. During the first pass a preliminarysegmentation is made. This is then held constantduring warp estimation to allow separatewarp-estimates to be made for each phoneme.Both a regular recognition phase as well asK-Means grouping has been used in their region-basedextension to VTLNThe group-based warping method above relieson a two-pass strategy where a preliminaryfixed classification is used during warp factorestimation, which is then applied in a final recognitionphase. Initial recognition errors canultimately cause a warp to be selected thatmaximizes the likelihood of an erroneous identity.Application of this warping factor will then145
Page 1 and 2:
Department of LinguisticsProceeding
Page 3 and 4:
Proceedings, FONETIK 2009, Dept. of
Page 5 and 6:
Page 7 and 8:
Page 9 and 10:
Page 11 and 12:
Page 13 and 14:
Page 15 and 16:
Page 17 and 18:
Page 19 and 20:
Page 21 and 22:
Page 23 and 24:
Page 25 and 26:
Page 27 and 28:
Page 29 and 30:
Page 31 and 32:
Page 33 and 34:
Page 35 and 36:
Page 37 and 38:
Page 39 and 40:
Page 41 and 42:
Proceedings, FOETIK 2009, Dept. of
Page 43 and 44:
Page 45 and 46:
Page 47 and 48:
Page 49 and 50:
Page 51 and 52:
Page 53 and 54:
Page 55 and 56:
Page 57 and 58:
Page 59 and 60:
Page 61 and 62:
Page 63 and 64:
Page 65 and 66:
Page 67 and 68:
Page 69 and 70:
Page 71 and 72:
Page 73 and 74:
Page 75 and 76:
Page 77 and 78:
Page 79 and 80:
Page 81 and 82:
Page 83 and 84:
Page 85 and 86:
Page 87 and 88:
Page 89 and 90:
Page 91 and 92:
Page 93 and 94: Proceedings, FONETIK 2009, Dept. of
Page 127 and 128: Proceedings, FOETIK 2009, Dept. of
Page 143: Proceedings, FONETIK 2009, Dept. of
Page 195 and 196:
Page 197 and 198:
Page 199 and 200:
Page 201 and 202:
Page 203 and 204:
Page 205 and 206:
Page 207 and 208:
Page 209 and 210:
Page 211 and 212:
Page 213 and 214:
Page 215 and 216:
Page 217 and 218:
Page 219 and 220:
Page 221 and 222:
Page 223 and 224:
Page 225 and 226:
Page 227:
Department of LinguisticsPhonetics
show all

Proceedings Fonetik 2009 - Institutionen för lingvistik

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?