On parameter filtering in continuous subword-unit-based speech recognition

ON PARAMETER FILTERING IN CONTINUOUS SUBWORD-UNIT-BASED SPEECH RECOGNITION Pau Pachis-Leal and Climent Nadeu Universitat Politkcnica de Catalunya Barcelona, Spain e-mail: paupag @ gps.tsc.upc.es ABSTRACT Simple IIR or FIR film have bem widely used in isolated or COMCC~C~ word recognition tasks to filter the time sequence of speech spectral parametas, since, despite their simplicity, they significantly improve recognition performance. Those film, when applied to continuous speech recognition, when phonema sized modelling units w used, induce spectral transition spnading and a cross-boundary effect In this work, we show how the use of contextdependent units reduces the side effects of the filters and may result in improved recognition performance. When dynamic parameten are not used. filtering scans to be especially useful, even for clean speech, and when they are, film do well under unmatched training and testing conditions. 1. INTRODUCTION At present, IIR or FZR filtering of the time trajectories of sp&ch speceal parameters is widely used since. albeit conceptually simple and amenable to a real time implementation, it allows a significant improvcwnt in recognition performaace In a logiuithmic spectral domain, whm a convolutive distortion time trajectories helps to alleviate the linear distortion due to the be bandpass, ratha than higbpass. Therefore, they can do somcthingelsethandycan~gtheDCcomponent Tbe study of the average long-tam power spec” [l] of the time sequences of spectral parameten, denoted by T(@, where 8. oh rcfd to as modulation frequency, is the fnquency countapar~ of the frame index n, shows that such filters, like those which an used to provide dynamic features or supplrmentary parametas, actually have two components: a differrntiation, oh implemented with a zero at or close to SI. which attenuates the low frequency region and approximately equalizes the nst of the long-tam specr”, and a lowpass component, which discards the high frequency zone of the long-tam spectrum, whcre the estimation mr variance is greatest and which has ban unduly enbanad by the first component [I I. With one feature, substitutive parameta filtering has also been shown to yield a substantial improvement for clean speech, although in this case the linea distortion due to the acoustic channel is fix4 through an enhancement of the time dynamics of the time trajectories of speech fawe vectors and through a reduction of speaker variability [21. becomes additive, filtering out the vay low frequencies of the SlOW-naying acoUStic channel. HOWCV~. most film rued tatd to However, those film have most often been applied to isolated or connected word recognition tasks [2,3], when the average word length is far higher than the effective length of the impulse response of the film. since the latter induce a spreading of spcctnl transitions and make the cumnt analysis output depend on its neighbouring context. Such side effens are critical in continuous subword-unit-based speech recognizers, for which filters may worsen recognition pafonuance [4]. Contextdependent subword units are frequently used in this frameworli since they can model speech coarticulation. In this work, we show how, as suggested by [5], the use of contextdependent units tackles the problem posed by the cross-boundary effect brought about by the filtering and may make the latter advanugcuus. we apply several fltcrs to the paramctuised utterances of two continuous speech databases and for diffaent sets of contextdependent units and number of features (one or two, that is to say, with or without addition of dynamic features or supplementary parametas). By doing so, in a nDmkc of cases. we get a twofold improvement in recognition performance: the one provided by the use of contexf-deptodent units and the additional improvement supplied by 6ltaing. ’Ihis file is extraordinariy simple, and often helps to further reduce the recognition error rate signiiicaarly. 2. RESULTS WITH ONE FEATURE Two different databases were used for phonetic classification wcpaimsu: the Spanish EUROh4.1 database 161, hereafta referred to as DB1, aad the Spanish ScntencesUPV database (which was recorded by the Univasitat Polit&nica de Valhxia), we will call it DB2. Both are used for speaker-independent tcsts. which 186 are different) from 43 speakers for Oaining, and around 12500 PLUS (in 225 uttaances, 61 of them different) from 17 speakers for resting. The laaa has 21667 PLUS (in 839 unaances, 120 being unique) fium 7 speakers for training and 5610 PLU’s (150 uacrauce~, 50 different) from 3 spealras for cepsaal coefficimts 0 plus an energy coefficient. 25-ms analysis windows with a 10-ms analysis step were used. Each of the PLU’s, as well as the silence model, is modelled by a threG state left-to-right continuous obsavation density HMM. The ncogniza was the HMM toolkit m). with three Gaussian mixtures pa state. The contartdepondmt labelling scheme is the same as m al. As the number of context-dependent phones (CDP) The formcr comprises nearly 37000 PLUS (in 842 utterances, of testing. Both databases are pa” ‘sed with 12 Mel-frcqu~~~y 1065 is too high, both in theory and in this particular database [T, and no statistically reliable model can be obtained for most of them, a threshold for the CDPs is defined which sets the minimal number of training tokens for the corresponding CDP model to be trained. If a CDP is not frequent enough (i.e. if the number of times it appears in the training corpus text does not attain or surpass the threshold), then it is the corresponding CI (context-independent) model which is trained. The coverage rate for a set of contextdependent units is defined as the fraction of the (training or testing) corpus text that can be labelled using this set. In this work, we have only used right context dependent phones (RCP), triphones (TrP) or a combination of both. The training co~pus coverage rate and the number of units for each of these sets is shown in Table I. I RCP 35 I 194 1 94.2% I Table 1: Analysis of the different sets of CD units. All the sets include the 26 CI phones to attain a 10% coverage rate. The filters we have used an (lowpass) Slepian filters [I] prexded by a hghpw) equalizing ater with transfa haion H(z)=I-d, where r is close to 1. Slepian filters have fw~ pamneta, the length L and the bandwidth W, which if catain conditions hold, gives the upper cutoff frequency (bandwidth) of the film. A Slepian filter of length L and bandwidth W is denoted as (r, W). We have also used the classical firstorder derivative window, which amounts to a FIR filter of order 5, and refer to it as regression. Simple W filters, the so-called RASTA filters, with a single complex pole and a numerator identical to the regression aforementioned filters substitutes for the original feaant, with no filter, have also bem tested In this &on, any of the three addition of supplementary parametas. 2.1. Results with DB1 (clean speech) Figure 1 shows the recognition results for DBl and one feature. We see that filtcriug somewhat worsens the recognition results for CI phones. The further we move to the right, the higher the improvement of the error rate (iomdly defined as 100%- %accuracy) with respect to the non-6ltcrcd case. For CI phones, this improve“ equals -0.97% (the best ater being the Slepian (5, 32)). whereas for the last Wrperimnt it is 11.4% (the regression filter is the best). Using CD phones allows us to have an 18.4% improvement in error rate (for the non-filtered case, and from the first to the last urpairmnt) and filtering enables us to further duce the error rate (by another 11.4%), although clean speech (acquired with a slngle microphone with little background noise) has been used. As was expected for continuous speech [l], filters of length greater than 5 were found to give wone results. Figure 1: Recognition results (%accuracy) for DB1 and one feature. 2.2. Results with Unmatched Training and Testing Conditions Fip 2 shows recognition results when the whole of the DB1 database is ustd for training and the whole of DB2 is used for testing. Those two databases were recorded with a different microphone and have some difkmces in dialect, speaking rate, etc.. In such conditions, CD units do not improve so much with respect to the CI case (7% for the non-filtered case), since other limiting factors than coarticulation, such as the different microphones. hold. Fdtuing. however, rounds off the work done by Q) units by providing a fuaher 36% improvement, which equals 24% only if no CD units are used. Figure 2 Recognition results (%accuracy) for the mismatched case @Bl used for training, and DB2 for testing). 23. conclusion Conten-dcpcndcnt units manage to take into accouIlt the natural dependence (caused by coarticulation) between phonermsizui units. It appears that they also succeed in incorporating the filterinduced dependence into the model, since, both for clean speech and for heavy mismatched conditions, filtering is the most advantageous for sets of context-depdent units with as many triphones as possible and triphones model a broad temporal input 1066 context similar to the extension of the tilm we have used. ?he reason for the success of CDPLUs seems to lie in the selection of contexts which they perform With contextdependent units, the contexts which an pooled together to train a model are more homogeneous than with context-independent units and thus are more robust to the transition s@ng produced by the filter. That increased robustness allows to take advantage of the beneficial effects of the filter such as alleviation of the effect of the acoustic channel or selection of the frequency zone where the discrimination capability of the spccch units is grcatcst However, we have worked so far with just one feature. In most practical recognition systems, two features at least arc used for increased performance, so the experiments carried out for one feature should be extended to a higher numbcr of features. The next section reports experiments using two features. 3. RESULTS WITH TWO FEATURES For the two-feanue we, the regression filter of length 5 (firstorder derivative window) presented above is used to supplement the first feature. In facs two strategies have been studied: the first, haeafter referred to as series strategy, applies this filter to the first filtered feaaue, where the filter is a short FIR filter (a Slepian of length varying between 5 and 9) or an W filter (a RASTA with a pole at .z=r). whmas the second strarcgy, the parallel strategy, applies the ngression filter to the original non-tilted feature to give the final second feature, while the hm (as is the case for the series strategy) is the original feature filtered with a FIR filter or an W filter. 3.1. Results with DB1 for the Series Strategy I I 1 No Filter I 55.15 I 65.15 I In this case. filtering seems detrimental to recognition more than attenuating the low fnsuency region, yields good Table 2: Recognition results @Bl, %kgcc) for the series shatcgy pcrfomance and contcxt~dcnt anits se~n unable to cope with it. Only the W filter with a pole at -0.97, which does hale results. The long-term spectntm (mer@ ova all cepstral coefficients) for the non-filmed case (the first feature for the last row of table 2) can be seen in fip 3, and that for the IIR filter with -0.97, by itself and convolved with the impulse response of the repsion tilter. is shown in figure 4. Only the low frequency region of the spectra is shown. As can be sem from figure 4, the two filtered cases. which correspond to the first and second feahlre of the last but one row of table 2, only differ in tbe low frrquency region: both annul the DC component but aftawards the first (the W by itself) presents a sharp peak vay near OHz, while the second (me XIR followed by the ngrtJsion film) gocs up smoothly f” the zero at OHz to a rounded peak near 6 Hz. The W with =0.9 shows the same pattun, except that the two iongterm spectra arc lus different (the initial peak being less sharp). Similar plots to those of figure 3 for Slepian filters show that the long-term spectra of the file by itself and that of the filter convolved with the impulse response of the regression filter are too much alike. The redundancy in the 6nquency domain explains the mediocre rrsults, sixnilar to those with one fatuns The filter which works best and outdoes even the non-filmed case (for the CI experiment) is the only one where there is some complementariness in the long-tam spectrum even if it is so in a small yet perceptually important low frequency region [2]. The W filter with kO.9 gives results between those of the W fdter with ~0.97 and those of the RR filters; in fact, the simple IIR filters outperfom any FIR filter from k0.8. In oder to try to obtain more different spectra, the second strategy was adopted. -1 5 -20 B H.q*ncy Flgure 3: average long-tam spec” of DB 1 for the non-filtered case. 44 -%12s4 res10 &w; Figure 4: average long-tam spectnrm of DB 1 for W -0.97 by itself and followed by a regression 6lter. 1067 3.2. Results with DB1 for the Parallel Strategy As can be seen from table 3, the parallel strategy is not effective either and it seems pointless to filter (with or without contextdependent units) when dynamic parameten are included in the fmture vector, since the filtered feature (or dynamic feature) already improves the discrimination by filering the first featwe, and since both are handed over to the recognizer. Only removing the DC component (IIR r=O.97) seems worthwhile. Plots similar to figure 4 show that the long-term spectrum of the second feature for the parallel strategy resembles the first feature filtered with a FIR more closely than does the second feature for the series strategy, which agrees with the fact that FIR filters give worse results in table 3 than in table 2. For the W filters. the two long-term spectra for the second feature in both cases look very much alike, which explains the similarity of the results obtained. Tabie 3: Recognition results (DB 1, %cacc.) for the parallel strategy In order to determine whether filtering was useful when using two features, we carried out the experiments of seaion 2.2 (mismatch) using two features. 33. Results with Unmatched Training and Testing Conditions, Parallel Strategy Those results are given in table 4. In this case it is worth filraing: filtas improve recognition in all cases and the improvemeat is higher when context dependent units are used. Even FIR filters, indfective in the parallel strategy due to the redundance in the bands of the long-tenn spearum for DB1, work well. The &t for Re35 is better than that for Tr/RCP35 in ~ccuracy but not in comctness. The result has been kept as it is for the sake of consistency with the other results reported in this work lIRr-0.97 I 61.26 I 65.88 1 62.44 Table 4: Results (%accuracy) for the parallel strategy in the mismatch case. NoFilter I 48.16 I 50.90 I 51.91 3.4. Conclusion For clean speech without mismatch and using two feaxures, at least when the second is obtained by Bering with a regression filter, filtering anything else than the zero frequency component seems to have no advantages. Therefore the use of context-depeadent units only provides a better modelling of coaniculation, rather than allowing to take advantage of the benefits supplied by filtering (since there are no such advantages if the WO features are Considered). In mismatched training and testing, fltering does seem advantageous and context-dependent modelling does increase its efficiency, much like what happened with just one feature. 4. CONCLUSION Simple W and FIR filters have been applied to continuous speech recognition. If context-dependent units are used, the side effects of the filters may be remedied and their benefits taken advantage of, which results in a significant improvement in recognition performance that adds up to the one provided by the use of context dependent units. When no dynamic pamneters are used, filtering and CD units yield a substantial improvement even for clean speech. When conventional dynamic parameters are incorporated into the feature vector, filtering and CD units seem useful especially for unmatched training and testing conditions. CD modelling must be vay effective for improving performance of continuous speech recognition over telephone lines, where filters are usually used. In order to make filtering and CD units useful under aIl conditions, alternative dynamic features should be found 5.REFERENcEs 1. Nab, C, and Juang, B.H., Titering of Spectral Parametas for Speech Recognition”. Proc ICSLP’94 2. Nadeu, C., Pachh-Leal, P.. and Juang. B.H., “Fiitering the Tm Sequence of Spectral Para“ for Speakerlndependent CDHMM Word Recognition”, Pra. kospeech’95 (Madrid), pp. 923-26,1995. 3. Hanson, BA., and Applebaum, T.H., “Subband or Cepstral Domain Filtering for Recogninon of Lombard (Yokohama), pp. 1927-30,1994. and channel-Distorted speech”, Proc. IcASSP’93 -h), p~. II-70, II-73. 4. Chang, J., and Zue, V., “A Study of Speoch Recognition System Robustness to Microphone Variations: Experiments in Phonetic Classification”, Proc. ICUP’94 5. Hamansky, H., and Morgan, N., “R4STA Processing of Speech”, IEEE Trans. on Speech and Audio Processing, VoL2, No. 4, Oaober 1994. 6. ESPRlT Project Speech Technology Assessment in (Yokohama), p~. 995-98,1994. Multilingual application^ (SAM-A). Domat SAM- N6002,1993 7. Bonafontc, A, Estany, R., and Viva, E., “Study of Subword Units for Spanish Speech Recognition”, Prof. E’urospecch’P5 (Madrid), pp- 1607-10,1995. 1068