High-resolution speech signal reconstruction in Wireless Sensor Networks

Transcription

1 High-resolutio speech sigal recostructio i Wireless Sesor Networks Adria Pazarloglou, Radu Stoleru, Ricardo Gutierrez-Osua Departmet of Computer Sciece, Texas A&M Uiversity {adria, stoleru, rgutier}@cs.tamu.edu Abstract Data streamig is a emergig class of applicatios for sesor etworks that has very high badwidth ad processig power requiremets. I this paper, a ew approach for speech data streamig is proposed, which is based o a distributed scheme. This scheme focuses o balacig the eergy cosumptio amog odes i a sesor etwork by allowig lowresolutio streams from multiple odes to be fused at a cetral processig ode i order to produce a ehaced resolutio speech sigal. Simulatios ad experimetal results with real microphoe sigals are preseted. I. INTRODUCTION With the recet growth of Wireless Sesor Networks (WSNs), may advaced applicatio areas have received sigificat attetio. Apart from the traditioal low data rate applicatios implemeted i sesor etworks, a eed of supportig higher data rate applicatios, such as audio ad video streamig, has appeared lately. Specifically, there is a lot of iterest i real-time streamig for military surveillace purposes ad for emergecy situatios []. The lack of badwidth ad limited samplig rate capabilities of sesors reder the implemetatio of these applicatios challegig. Amog the requiremets of streamig applicatios is the high volume of data that eeds to be set to the sik. The basic drawback of obtaiig data from oe particular ode every time is that it could lead to a over-utilizatio of the radio of specific odes. The radio is oe of the most eergyiefficiet compoet o a sesor ad power supply is limited ad battery rechargig is almost iexistet. This evetually leads to depletio of the power supply of those odes sooer tha the other oes ad reduces sigificatly the etwork lifetime. Fig.. Streamig schemes with (a) oe ad (b) multiple odes listeig ad trasmittig data Aother factor to take ito cosideratio is that, i WSNs, the data from eighborig odes is usually highly correlated. Cosequetly, the trasmissio of the gathered data of those odes leads to redudat iformatio. Our aim is to desig a distributed scheme for data acquisitio that balaces eergy cosumptio while exploitig spatial correlatios betwee eighborig odes. I the process, we also reduce the badwidth ad processig requiremets for sesor odes. The cotributios of the paper are the presetatio of a ovel approach for high resolutio speech sigal recostructio ad the demostratio of the proposed approach, through performace evaluatio i simulatios ad a proof of cocept hardware implemetatio. The paper is orgaized as follows. Sectio 2 presets the related work. Sectio 3 details the desig of the proposed scheme for high-resolutio sigal recostructio. Sectio 4 presets ad discusses performace evaluatio results. Coclusios ad future work plas are give i sectio 5. II. RELATED WORK Curret approaches for streamig voice data i WSNs, assig the resposibility of trasmittig the data to a ode that is closer to the source of the evet, as i Figure (a). Magharam et al. [] implemets a etworkig platform i order to support real-time voice streamig. Aother attempt for streamig service support i order to support military surveillace applicatios is described i [2]. Commo to both studies is the eed to obtai uderstadable speech from WSNs with eergy cosumptio costraits. Thus, microphoe sample rates are lower tha the ormal 8kHz ad also, Adaptive Differetial Pulse Code Modulatio (ADPCM) is used to ecode the data ad thus reduce the trasmissio data rate. A high data-rate capturig system for sesor etworks has bee preseted by Greestei et al [3]. I this approach useraccessible devices (each capable of performig sophisticated sigal processig) ca tue the sigal processig parameters of the sesors i order to reduce the amout of data to be trasmitted. III. SPEECH SIGNAL RECONSTRUCTION FROM ASYNCHRONOUS DATA SOURCES I our proposed scheme, the odes that sese a evet sample with low data rates ad sed the data through differet routig paths to the sik, as show i Figure (b). The sik receives the udersampled data from differet odes ad recostructs a high-resolutio versio of the origial sigal (Figure 2) /9/$ IEEE

2 Fig. 2. Recostructio scheme The major advatages of this scheme are that it distributes eergy cosumptio i more tha oe ode, thereby icreasig the efficiecy ad the survivability of the WSN, ad also shifts processig requiremets from the odes to the sik. At the same time, this scheme does ot impose ay demads for tight time sychroizatio, ad istead exploits radomess i samplig times for each ode to recostruct a higher-resolutio sigal. O the cotrary, if the received sigals from multiple odes were to be sychroized, recostructio of a better quality sigal would be ufeasible. Fortuately, this sceario is highly ulikely, sice timig sychroizatio is really difficult to achieve. A. Desig of High-resolutio Sigal Recostructio I our approach, a optimal recostructio [4] is ot required, but rather a acceptable recostructio. A acceptable soud sigal is a sigal that has some predefied qualitative characteristics, such as deliverig a soud sigal that is accurate eough to be uderstood by a listeer. Give that the frequecy rage of a speech system is from 3 Hz to 35 Hz, where the majority of sigal eergy beig betwee 5 Hz to 25 Hz, samplig rates used i existig applicatios are 4kHz[]ad2kHz[2].However,thesampligrateiour proposed system ca be much lower depedig o the umber of the odes that liste to the evet at the same time. Our recostructio approach is based o the fact that streams from multiple sources with lower quality represetatios of the origial sigal provide differet pieces of iformatio about the origial sigal, provided that their samplig times are radomly shifted. Figure 3 illustrates the samplig ad recostructio. The origial sigal is deoted by x(t) ad the discrete-time sigals s i (k),i =,..., are oisy versios of x(t) produced i each ode after samplig (where k deotes the time idices at sesor ode samplig rate f s ad k deotes the time idices for the higher resolutio sigals at samplig rate f s = f s r). Sice the more efficiet compressio techiques (with compressio ratio 2: or better) such as Mixed-Excitatio Liear Predictive (MELP) ad Liear Predictive Codig (LPC) are too complex for sesor odes, we choose to compress s i (k) with 4-bit ADPCM which provides a compressio ratio of 4:. O the recostructio part, the received streams are first ADPCM-decoded ad the aliged i time before beig fused. Sice time delays betwee sigals from differet odes might Fig. 3. Distributed Data Streamig ad Sigal Recostructio ot be iteger multiples of their samplig rate, the sigals are upsampled to a commo samplig rate through iterpolatio. Time delay estimatio is performed from the resultig sigals s (k ) i order to alig them properly o the time domai. The output sigal is the formed by superimposig the aliged sigals ad a deoisig filter is applied. B. Time-Delay Estimatio Here, the term time-delay estimatio refers to the relative Time Differece Of Arrival (TDOA) betwee sigals received at spatially separated sesors. Time-delay estimatio is a commo research topic i sigal processig, where it is used as a first stage i applicatios such as speech ehacemet ad recogitio, ad source localizatio ad trackig. The most commo method for the time delay estimatio, betwee two waveforms x ad x 2, is the Geeralized Cross- Correlatio (GCC) fuctio [5]: where ˆd GCC (m) = N k= W [k]g xx 2 [k]e j2πmk/n G xx 2 [k] =E[X [k]x 2 [k]] is the cross-spectrum of x ad x 2, X i [k] is the discrete Fourier trasform (DFT) of x i (), N is the legth of the DFT ad W [k] deotes the weightig fuctio or prefilter. I the case of a costat weightig fuctio W [k], the GCC fuctio becomes a frequecy domai implemetatio of the CC fuctio (o-weighted cross correlatio). Differet solutios have bee proposed to choose a suitable weightig fuctio [5]. For the case of TDE i multichael cases, oe approach is to choose oe chael as a referece ad the TDOA for the rest of the chaels ca be estimated usig a TDE method. More robust estimatios ca be obtaied by exploitig the fact that more chaels ca be used as referece i a way to overcome errors caused by reverberatio ad oise. C. Recostructio from Low Resolutio Sigals We cosidered two approaches for the superpositio of the aliged sigals. I the first approach, we iterpolated the low 2

3 resolutio sigals ad the average betwee them was take x(k )= s (k d ) N where d is the delay for the sigal s. I the secod approach, we icreased the sigal samplig rate with upsamplig (e.g., fill the spaces betwee samples with zeros) as { s (k s (k), if k = k r, k =, ±, ±2,.. )=, otherwise ad the sigals s (k ) aresummedupas: x(k )= s (k d ) ad zero samples were fially replaced through liear iterpolatio. D. Noise Suppressio The oise suppressio stage is the last stage i the sigal recostructio process, ad is used to smooth the sigal ad remove possible backgroud oise. The use of multiple microphoes i observatios of the same evet, ca give a better idea of the source sigal ad it has bee a active area of research for the past decades. Noise suppressio methods used i microphoe arrays ( [6], [7]) for speech recogitio ad ehacemet are based o Wieer post-filterig techiques. These methods are applied i the fial recostructio stage. However, i high oise eviromets, sigals from multiple sesors ca be ehaced so that subsequet steps (i.e., samplig rate icrease, TDE) will perform better. A wide rage of oise suppressio techiques have bee proposed. Amog them the most popular are spectral methods, such as spectral substractio ad Kalma filterig. IV. PERFORMANCE EVALUATION I this sectio we preset performace evaluatio results for the recostructio of a origial speech sigal from multiple low resolutio sigals. This evaluatio icludes both simulatio ad actual recordigs. We perform simulatio of the recostructio scheme usig speech sigals already acquired ad preset i simulator libraries, as well as usig speech sigals obtaied i real laboratory experimets. To evaluate the quality of a recostructed sigal, we measure the Normalized Root Mea Square Error (NRMSE) [8] ad the PSNR (Peak Sigal-to-Noise Ratio). The NRMSE is give by: NRMSE = (x ˆx ) 2 E[(x µ(x )) 2 ] where x is the origial sigal, ˆx is the recostructed sigal ad µ deotes the mea value of the sigal. The PSNR is give by: PSNR = log Nx 2 peak x ˆx 2, where x is the origial sigal, ˆx is the recostructed sigal, x peak deotes the maximum value of x ad x ˆx 2 is the eergy of the differece betwee the origial ad recostructed sigal. A. Simulatios I order to establish proof-of-cocept, we first performed simulatios. A speech sigal sampled at 22 khz sample rate ad 6-bits per sample was used for the experimet. We produced a umber of udersampled versios of the origial sigal (as would have bee acquired by a set of sesor odes), with radom delays. Fially, we added Additive White Gaussia Noise (AWGN) at differet SNR levels. ) Accuracy of Time Delay Estimatio: Time-delay estimatio was implemeted usig the o-weighted cross correlatio. Low Sigal to Noise Ratio (SNR) ad the samplig rate icrease method ca chage the performace of TDE. Figure 4 illustrates how the level of oise ad differet sample rates affect the result of the delay estimatio. A sigal havig samples delay from the origial 22 khz sigal was created. Additive White Gaussia Noise (AWGN) was added ad both sigals were dowsampled with a rate of 5 ad. I the ext stage their samplig rate was icreased to reach 22 khz usig iterpolatio. Time delay estimatio for iterpolated sigals was accurate for SNR over 5 db. Delay Estimatio Error (samples) kHz 2.2kHz 22.2kHz Fig. 4. Delay estimatio error uder differet samplig rates ad oise levels after iterpolatio ( iteratios). Our experimets demostrate that the o-weighted cross correlatio usig iterpolated sigals gives accurate time-delay estimates for high oise levels ad a precisio of sample o average for SNR over 8 db. 2) Quality of the Sigal Recostructio: Figure 5 shows a schematic of the process that was followed i the simulatios. The time delay betwee each sigal versio that is created ad the referece, S k, is expressed with z Di, where i =2,...,. Noise suppressio was performed o the udersampled sigal represetatios usig a bidirectioal ostatioary Kalma filter. Figures 6 show the recostructio quality i terms of the NRMSE ad the PSNR, uder differet oise levels. The two speech dft.wav with utteraces from a male adult, as provided by the MATLAB library 3

4 samples cotribute equally with the iterpolated oes. Figure 7 plots these sigals (the first /6 portio of speech dft.wav is plotted) for a SNR of 35 db, as well as the clea origial sigal. Origial sigal Fig. 5. Schematic for the simulatios.5 Iterpolated sigal from oe ode approaches are applied for the recostructio of a 7.35 khz speech sigal from five oisy udersampled versios ad compared agaist the speech sigal produced after iterpolatio of oe udersampled versio. For this simulatio, ADPCM ecodig was ot applied Recostructio from iterpolated sigals NRMSE Recostructio uder differet oise levels usig deoised data data from 5 odes usig uoised data from 5 odes usig deoised iterpolated data from 5 odes usig uoised iterpolated data from 5 odes usig deoised iterpolated data from oe ode Reoctructio from upsampled sigals Fig. 7. Sigal plots (35 db SNR) Recostructio uder differet oise levels Figure 8 shows the samplig rate i Hz for each ode, i accordace with the recostructed sigal quality. AWGN was added o the udersampled data to give a SNR of 35 db Recostructio performace for differet samplig rates data from oe ode data from 3 odes data from 2 odes data from 5 odes PSNR 2 NRMSE usig deoised upsampled data from 5 odes usig uoised upsampled data from 5 odes usig deoised iterpolated data from 5 odes usig uoised iterpolated data from 5 odes usig deoised iterpolated data from oe ode Samplig rate (Hz) i each ode Fig. 6. Recostructio quality (NRMSE, PSNR) usig data from ad 5 odes uder differet oise levels. Oe recostructio approach is to take the average of the iterpolated sigals, while the other is to take the sum of the upsampled sigals ad fill the zero samples usig iterpolatio. Ituitively, the lower quality recostructio usig iterpolated sigals, ca be justified from the fact that the actual Fig. 8. Recostructio quality usig oisy data (SNR 35 db) from 5, 3, 2, or ode i order to recostruct a speech sigal with 225 Hz sample rate. B. Experimetal Results Experimets with real recordigs were performed i a computer lab. Speech sigal acquisitio was made from six 4

5 differet positios placed i a straight lie ad havig cm distace betwee them. Oe computer speaker was placed at a distace of 5 cm of the lie ceter. The recorded sigal had 225 Hz sample rate with 6-bit samples ad oe chael. Timig delays were estimated usig the six origial sigals, which were give as iput to a o-weighted cross-correlatio fuctio; oe of the sigals was selected as a referece. The six aliged microphoe sigals were dowsampled by a factor of 2 to create 2.5 Hz sigals. The low rate sigals were upsampled by a factor of, superimposed ad iterpolated to fill the zero samples. As a result, the output sigal had a samplig rate of 25 Hz. NRMSE ad PSNR were measured from a origial microphoe sigal dowsampled by a factor of 2, ad were: ad 33. respectively for the recostructed sigal ad for a.25 Hz sigal that had its rate icreased usig iterpolatio by a factor of 9.6 ad for a 5.5 khz sigal that had its rate icreased usig iterpolatio by a factor of 2 The above results show that it is feasible to obtai a clear sigal after combiig multiple lower-rate sigals with o uderstadable cotet (such as. khz speech sigals). V. CONCLUSION We have outlied a ew approach for speech data streamig based o a distributed scheme, that targets a balaced eergy cosumptio amog odes i a sesor etwork. Amog the advatages of this scheme is that it does ot impose further processig or sychroizatio requiremets for the sesor odes. Moreover, the method takes advatage of the radomess of the samplig of each ode. Aother iterestig aspect would be to implemet a feedback algorithm that could chage the recordig phase of oe or more odes so as to have a more uiform distributio of the samples withi a samplig period. This approach opes a whole ew field for research. A methodology must be developed to defie which odes should start recordig based o their locatios, takig ito accout the topology of the WSN. A more dese etwork could lead to more distributed samplig, leadig to overall better etwork survivability. I the experimets, the recostructio method does ot take ito accout the reliability of the sigal itself. Samples that are raw data should be cosidered as more importat i the recostructio tha samples that result of a iterpolatio fuctio. This could lead to better quality i the recostructed sigal. I additio, further research is required to fid which TDE method is more appropriate to use for low rate speech sigals ad for specific eviromet coditios (i.e. reverberatio, oise), sice the o-weighted cross-correlatio method is ot robust whe it accepts reverberat sigals as iput. We discussed ad evaluated two recostructio approaches, oe from upsampled sigals ad oe from iterpolated sigals. I a real implemetatio, a hybrid solutio that combies the best of both of them should be cosidered ad evaluated, such as a solutio that uses iterpolated sigals with weights. Fially, the proposed approach has iterestig implicatios for security-related applicatios. A multipath scheme where very low rate sigals are circulated towards a sik could make eavesdroppig cosiderably difficult, provided that lowresolutio speech sigals from each microphoe are ot uderstadable by a listeer. Our immediate goals are to evaluate the efficacy of our approach by developig a real sesor etwork implemetatio. I additio, we will ivestigate multipath routig i mobile eviromets ad additioal filterig schemes to ehace sigal recostructios. REFERENCES [] R. Magharam, A. Rowe, R. Rajkumar, ad R. Suzuki, Voice over sesor etworks, i RTSS 6: Proceedigs of the 27th IEEE Iteratioal Real-Time Systems Symposium. Washigto, DC, USA: IEEE Computer Society, 26, pp [2] J. Zhag, G. Zhou, S. H. So, ad J. A. Stakovic, Ears o the groud: A acoustic streamig service i wireless sesor etworks, i IPSN 6, 26. [3] B. Greestei, C. Mar, A. Pesterev, S. Farshchi, E. Kohler, J. Judy, ad D. Estri, Capturig high-frequecy pheomea usig a badwidthlimited sesor etwork, i SeSys 6: Proceedigs of the 4th iteratioal coferece o Embedded etworked sesor systems. New York, NY, USA: ACM, 26, pp [4] O. Jahromi ad P. Aarabi, Theory ad desig of multirate sesor arrays, Sigal Processig, IEEE Trasactios o, vol. 53, o. 5, pp , May 25. [5] C. Kapp ad G. Carter, The geeralized correlatio method for estimatio of time delay, Acoustics, Speech ad Sigal Processig, IEEE Trasactios o, vol. 24, o. 4, pp , Aug 976. [6] R. Zeliski, A microphoe array with adaptive post-filterig for oise reductio i reverberat rooms, Iteratioal Coferece o Acoustics, Speech, Sigal Processig (ICASSP), vol. 5, pp , Nov [7] I. McCowa ad H. Bourlard, Microphoe array post-filter based o oise field coherece, Speech ad Audio Processig, IEEE Trasactios o, vol., o. 6, pp , Nov. 23. [8] E.-B. Fgee, W. Phillips, ad W. Robertso, Comparig audio compressio usig wavelets with other audio compressio schemes, Electrical ad Computer Egieerig, 999 IEEE Caadia Coferece o, vol.2, pp vol.2,