Improving the Sound Recording Quality of Wireless Sensors Using Automatic Gain Control Methods

Transcription

1 BULETINUL ŞTIINŢIFIC al Universiăţii POLITEHNICA din Timişoara, România, Seria AUTOMATICĂ ŞI CALCULATOARE SCIENTIFIC BULLETIN of The POLITEHNICA Universiy of Timişoara, Romania, Transacions on AUTOMATIC CONTROL and COMPUTER SCIENCE, Vol. 56 (70), No. 2, June 2011, ISSN X Improving he Sound Recording Qualiy of Wireless Sensors Using Auomaic Gain Conrol Mehods Gábor Goszolya * and László Tóh ** * Deparmen of Informaics, Universiy of Szeged, Árpád ér 1., 6720 Szeged, Hungary Phone: (+36 62) , Fax: (+36 62) , ggabor@inf.u-szeged.hu, WWW: hp:// ** Research Group on Arificial Inelligence of he Hungarian Academy of Sciences., Tisza Lajos kr. 103., 6720 Szeged, Hungary Phone: (+36 62) , Fax: (+36 62) , ohl@inf-u-szeged.hu, WWW: hp:// Absrac When performing speech recording i is desirable o have he speech signal in as a high qualiy as possible. In everyday recording condiions one of he mos imporan aspecs of sound qualiy is o have a uniform volume level, because i is very hard o undersand (and o auomaically recognize) an uerance wih a volume level ha varies considerably. Of course his uniform volume level should also be an average one, avoiding eiher oo loud or oo quie recordings. To overcome his problem usually an approach called Auomaic Gain Conrol is used, which is an adapive mehod for conrolling microphone sensiiviy (gain). Wireless sensors are recen, low-powered devices, which are for recording and ransmiing observaions such as speech, hus hey are a good area for applying auomaic gain conrol. Due o heir low power consumpion, however, only very simple soluions can be implemened. Here we will presen a general gain conrol algorihm, hen inroduce wo variaions ha we es in a siuaion which simulaes he acual use. We perform evaluaions by using wo ypes of measuremen: he firs one compares local volume levels o recordings made under condiions, while in he second we measure he undersandabiliy of he recordings made by applying sandard speech recogniion echniques. Our resuls in boh cases confirm ha i is indeed an area where auomaic gain conrol can be applied, and ha boh our algorihms perform well in pracice. Keywords: wireless sensors, sound qualiy, auomaic gain conrol, volume level, speech recogniion. I. INTRODUCTION Wireless sensors and wireless sensor neworks have become increasingly popular recenly. Their main applicaion is he monioring of heir environmen like movemen deecion and measuring emperaure and ligh. They are also capable of recording audio daa, which calls for poring a number of speech processing applicaions. Being a relaively new area, however, a number of quie basic issues for hese sensors have o be addressed. Because of heir limied processing capabiliies, even he poring of he mos basic algorihms migh require some modificaions. In his paper we will focus on a special problem of disan speech recording, namely he auomaic adjusmen of he recording gain. The above problem arises from he fac ha he posiioning of wireless sensors canno be known in advance. From he viewpoin of speech processing i means ha a sensor having a microphone and ransmiing wha i records does no know how far i is from he speaker. I could also be he case ha his arge is coninuously changing is posiion, i.e. moving away from or owards he sensor. Anoher siuaion migh be ha here are several speakers, each a a differen disance from he sensor; bu regardless hese difficulies, he wireless sensor should always ry o record he acual speech in as high a qualiy as possible. Perhaps he mos imporan facor affecing recording qualiy of a wireless sensor is ha of seing he gain of he microphone. This way he sensiiviy of he recording process can be affeced. A good auomaic gain conrol (AGC) algorihm [1] can eliminae or a leas smooh he above-menioned jumps in he volume level, which makes he effecive processing of he signal more sraighforward. The srucure of he paper is as follows. Firs we shall describe he problem of auomaic gain conrol and is imporance in our sudy. Then we will inroduce our algorihms applied and explain how hey work in deail. Nex, we will describe he evaluaion mehodologies used and our esing environmen. Finally, we will presen he resuls obained and draw our conclusions. II. AUTOMATIC GAIN CONTROL IN WIRELESS SENSORS The goal of auomaic gain conrol is o dynamically adjus he sensiiviy of he microphone, based on he acual acousic observaions. I is used in various siuaions where 47

2 signals having differen srenghs (usually from differen sources) are presen. A ypical example is ha of elephones or cell phones [1], bu i is also used for conrolling he ampliude of inerfering signals using lasers [2] or in pacemakers [3]. Every case is special, however, hus our sensor-based environmen also has is special requiremens, which should be considered when developing applicaions for i. Perhaps he mos imporan one is ha, as hese sensors were designed o have excepionally low power consumpion, hey have very limied resources: hey usually have a low-capaciy processor, and an exremely small amoun of RAM. They communicae via radio waves, which also have a limied bandwidh. In his scenario, applicaions designed o work on wireless sensors have o use as small a CPU ime and RAM as possible. The main reason for using gain conrol in wireless sensors is heir limied resoluion: for each sample we can represen i only using 10 bis of informaion, so we should aemp o use as much of i as possible. We could also follow he sraegy of sending he informaion recorded as is, and amplify i laer, in he device ha receives he speech daa. I would clearly have he advanage ha we would no be handicapped by he compuaional limiaions of sensors, so we could use CPU-demanding algorihms, or ones ha require more memory. In his scenario, however, here would be a clear loss of informaion as each sample has his fixed 10-bi long represenaion. For a louder-speaking person large-ampliude values would ge clipped, which could have been avoided if we had been using a lower gain value. In oher cases, when he speech recorded is oo quie, we would no be using he whole 10 bis available, bu only a par of i, which could be avoided by using a higher gain value. And i is clear ha in boh cases he qualiy of he recordings would suffer. Forunaely in our wireless sensor environmen he gain can be adjused. I is represened on one bye only: he higher his value is, he more sensiive he microphone becomes, and vice versa: a lower value means less sensiiviy. From experience we know ha he microphone has a quasi-linear sensiiviy as a funcion of he gain value, and is defaul value is 64. III. THE ALGORITHMS APPLIED Nex, we will describe our algorihms inroduced for auomaic gain conrol. An arbirary AGC mehod can be summed up in one senence: if he signal is oo srong, lower he gain; oherwise, if he signal is oo weak, increase he gain. This simpliciy, however, leads o difficulies a he poin of is applicaion: we could find no menion of a general AGC algorihm in he lieraure. I seems ha all he deails are heavily applicaion-dependen, which is convenien for deermining he hreshold values for he conceps oo srong and oo weak, or finding ou he frequency of performing he checks and modifying he gain. On he oher hand, here is no sandard way even for he process of increasing or decreasing he gain (which seems o be an issue for which quie general soluions should exis), and furher, we have o come up wih a way of measuring he srongness or weakness of he signal (loudness in he case of sound signals). In his paper we sough o inroduce acual algorihms for our paricular problem, hus all hese open issues have o be addressed. These issues were solved in a similar way as ha for our algorihms inroduced, and of course hey suffer from he same hardware limiaions, hus he resuling mehods are indeed quie similar. Due o his, we begin wih heir common properies, defining a general gain conrol algorihm. In he following we will denoe he acual gain value by gain, while gain' will be he new gain value; of course 0 gain, gain 255. A. Working wih Packes One characerisic of wireless sensors is ha hey communicae via radio waves, sending a small chunk of daa called a packe a a ime. In our case here can be a mos 114 byes in a packe; assigning one wo-bye ineger o he number of he packe (o be able o recognize missing packes) 112 byes remain. As he digial-analogue converer (DAC) of he microphone supplies observaions of 10 bis per sample, we can send bi samples in a packe, and he wo remaining byes could be used o send exra informaion. Of course he acual values may vary somewha beween differen hardware, bu he main concep (having small-sized packes which implies working wih a small, fixed number of samples a a ime) remains he same. This arrangemen makes i sraighforward o handle all observed daa in groups of 88 samples, i.e. A = a 1 a N, and in our archiecure N=88. Thus he inpu signal is organised ino a (heoreically) endless flow of packe A i -s, he las one being A ; i makes i plausible o perform he same acions for each packe A i. Of course some procedures may be performed afer every nh packe, bu he same lines of code can sill be execued for each packe, and hese do no refer o samples of anoher packe. (Alhough using some value represening anoher packe as a whole (like is energy level) is allowed.) B. Relying on he Energy Level Anoher common feaure of our algorihms is ha hey boh rely on he energy level of he speech signal observed. As he energy of a signal is closely conneced wih is volume level, conrolling he energy level means conrolling he volume. Moreover, he calculaion of energy is compuaionally very cheap, which is a vial requiremen in our case. In he acual soluion we followed he packeoriened sraegy described above. Firs we calculaed he mean of he values in he packe as he signals may conain a DC bias, i.e. 48

3 N 1 base A = a i. (1) N i = 1 Nex we calculaed he energy level of he packe, which is usually done by aking he squared sum of is values. Due o speed limiaions we did no raise hem o he second power; insead we jus added up he absolue values of he difference of he sample and he above-calculaed mean value, i.e. N energy = a base. (2) A i= 1 This value was hen reaed as he energy level of packe A, used boh for voice aciviy deecion (see below) and for measuring he loudness of he acual signal observed in order o conrol he gain. As our sensor boards produced sound signals wih a 8861 samples per second sampling rae, one packe corresponds o roughly 10ms of he speech signal. We sored he energy levels of he las 50 packes, examining abou half a second a a ime; he sum of hese values gave he full energy of his inerval, i.e. i = energy. (3) i= 49 Gain adjusmens (including hose of voice aciviy deecion) were made every 10 packes, i.e. in roughly 100ms ime inervals. C. Voice Aciviy Deecion A ypical gain conrol algorihm seeks o normalise he level of he observed signal o a pre-se value. This aim, however, becomes counerproducive if here is no voice aciviy a all; in his case only he basic noise of he microphone is presen, which will be amplified o he highes level available by seing he gain very high. When he silen period ends, he firs par of speech will be overamplified, leading o clipping and i will resul in a loss of informaion. To overcome his problem we should deec hese longer periods of silence, and se he gain level o an inermediae value here. This way we will no lose oo much informaion a he beginning of he nex speech porion (eiher if he speaker is oo close or oo far away), and hen we can reac o he curren volume level quie quickly. In he field of speech recogniion he problem of deecing longer silen pars is called Voice Aciviy Deecion, and here exis several algorihms o solve i [4] [5]. The simples of hem are based on calculaing some kind of loudness of he signal presen, and reaing i as silence if his loudness remains under a hreshold for a long period. We also followed his approach of voice aciviy deecion; hus, o use sparingly he limied resources of wireless sensors, we based he Voice Aciviy Deecion process on he same measuremens as he acual gain conrol, reusing he values calculaed above. Tha is, we used he full A Ai energy level of he las 50 packes (roughly 500ms), calculaed in (3). For he sake of simpliciy we assumed ha he gain value was he same hroughou his inerval, which in pracice worked quie well. Afer every en packes (i.e. every 100ms) we checked o see wheher he energy level for he las half a second divided by he gain value and a small consan sayed below a hreshold T SIL ; if so, we considered i silence, and se he gain o an average value gain SIL. So we used he condiion gain + c 0 < T where c 0 was deermined empirically. (I was necessary because using a gain value of 0 does no mean silence in our archiecure.) If we inerpreed he las par of signal as silence, i also mean ha we did no adjus he gain any furher his ime. D. The Common Pars in he Two Algorihms Nex we will inroduce wo algorihms which, based on heir observaions, dynamically adjus he gain level. Besides he similariies described above (working wih packes, using packe-size energy levels and applying voice aciviy deecion) here is anoher common aspec of heir behaviour. They seek o keep he oal energy of he las 500ms (i.e. ) a a T IDEAL level; for his, if he full energy is lower han a hreshold T LOW, he gain is increased, whereas if i is higher han T HIGH, i is decreased. T LOW was se o 80% of T IDEAL, while T HIGH was 120% of i, which values were deermined by simple preliminary ess. We used anoher hreshold (T UHIGH, being wice ha of T HIGH ): if he full energy level exceeded his value, we supposed ha he signal was clipped because is loudness was excessive, reaing i as an overdrive was presen, which called for an immediae and drasic reacion. A such imes he gain was decreased by using he formula gain' = ¾ gain based on preliminary ess again. For he pseudocode of his general algorihm, see Table 1. TABLE 1. General Gain Conrol Algorihm. Sep Insrucion 1 is he oal energy of he las 50 packes 2 if / (gain + c 0) < T SIL hen 3 gain is se o an inermediae value gain SIL 4 else if > T UHIGH hen 5 gain is reduced o is 3/4 h 6 else if > T HIGH hen 7 decrease gain 8 else if < T LOW hen 9 increase gain 10 end if E. Equal-Sepping Gain Conrol Algorihm In his algorihm (ES GCA) we adjus he gain in small, equal seps: o increase he gain level we used he formula gain' = gain + sep, SIL, (4) 49

4 and we applied gain' = gain sep o lower i. Naurally he bes value of sep has o be deermined, which mus be a posiive ineger. This algorihm is a simple and sraighforward way of conrolling gain. F. Weighed-Average Gain Conrol Algorihm The previous algorihm makes equal-sized seps up- and downwards regardless of he difference beween he measured energy level and he one. However, i migh make sense o have big seps if his difference is big, and only small ones if i is small. The second algorihm (he WA GCA) follows his sraegy. Assuming ha he energy level of a recording is linearly proporional o he gain used o obain i (which is roughly he case for our paricular sensor boards), we have ha gain =, (5) gain where is he full energy level recorded under condiions, deermined by preliminary ess, gain is he corresponding gain, is he oal energy level recorded, and gain is he acual gain value. This formula can be rearranged o express he expeced gain level; ha is, gain gain =. (6) Having defined he problem and presened he algorihms, nex we will urn o a descripion of he esing process. A. Hardware In his sudy we used Crossbow Iris sensor nodes (moes) ha have a 7.37 MHz processor wih a RAM of 8K byes and a programmable flash memory of 128K byes. The microphone and oher inpu peripherials are locaed on a piece of hardware ha can be aached o he moe, on he so-called sensor board. We had Crossbow MTS300 sensor boards, which, besides he microphone, also conain ligh and emperaure sensors. B. The Recording Environmen To simulae real-life condiions we made recordings a four differen disances: we posiioned he sensor a 20, 50, 100 and 200 cenimeers away from he speaker. The 50 cenimeer-long disance served as an recording posiion, 20 cenimeer as he speaker being oo close o he sensor, while he 100 and 200 cenimeer values simulaed speakers being quie far away. Tesing was performed on Hungarian broadcas news: a five-minue-long signal was used. I was hen modified o conain a wide range of volume level changes such as slowly growing quieer (simulaing he speaker going away from he microphone), slowly growing louder (he speaker is approaching he microphone) and sudden jumps (simulaing muliple speakers being presen a differen disances from he microphone). The same modified signal was played a differen disances o he sensors wih differen gain conrol algorihms, bu oherwise he siuaion (volume level, posiion) was exacly he same as before, so ha heir performances could be compared. Using his formula direcly, however, would lead o frequen big jumps in he gain level used, which would clearly harm he recorded speech signal. To couner his effec we averaged he previously used gain level wih his calculaed one weighed via gain ' = w gain + (1 w) gain, (7) where 0 w 1 is a weighing consan. Afer subsiuing gain from (6) we ge gain gain ' = ( w + ( 1 w ) ). Wih he w weighing facor we can make he ransiion of gain much smooher by eliminaing sudden jumps. Needless o say, he opimal value for w firs has o be deermined. IV. THE TESTING PROCESS (8) We made recordings using he wo gain conrol algorihms a all four disances. To ge a reference recording we recorded he original signal (i.e. he one wih consan volume) a 50 cenimeers wih a fixed gain value. We hen made wo addiional recordings a all four disances. Firs he original, consan volume-level signal was recorded wihou using gain conrol; hese (he basic recordings) were reaed as he ones in quasi- circumsances: he speaker is in a fixed posiion and has a consan voice level, bu i is no necessarily he opimal one: i could be oo high or oo low. The performance of he one a 50cm, compared o he reference recording, served as he glass ceiling: as hese recordings are as near idenical as possible, is score is he highes one available (a leas in heory), and his value canno be exceeded no maer wha sophisicaed gain conrol algorihms we develop. Nex he signal wih changes in volume level was recorded a each of he four disances, sill wihou gain conrol. These were he baseline recordings, and heir performance scores simulae he resuls we would ge in real-life siuaions (muliple or moving speakers) wihou gain conrol. For a lis of recordings made, see Table 2. 50

5 TABLE 2. The recordings made and he disances used. Disance Recording Volume 20cm 50cm 100cm 200cm Reference consan Basic consan Baseline varying ES AGC varying WA AGC varying C. Evaluaion via he Energy Level One sandard way of evaluaing a gain conrol algorihm is o calculae he energy levels of he reference and he resuling recordings, and ake heir difference or raio. Forunaely he energy level is very easy o calculae, and he volume of speech is direcly relaed o is energy, so he more similar he energy levels of he original and he gainconrolled recordings are, he closer hey are. The energy is calculaed by aking he squared sum of he samples, i.e. T 2 energy = s( i), (9) S i = 1 where s is he observed speech signal having a lengh of T. (Noe ha he evaluaion of he resuling signals was no done on he wireless sensors, so we could apply even compuaionally demanding algorihms for i. This way we were no forced o calculae he energy using absolue values, bu we could ake he squared sum of he samples insead.) In his form i has one value for he whole signal, which could indeed be of ineres; for his reason we will calculae energy S engraio S =, (10) energy where energy ref is he oal energy of he reference recording. This value, however, ignores he local variaions wihin he signals: wo recordings wih quie differen local values could have he same overall energy level. To overcome his weakness we inroduced anoher measure. We calculaed he energy levels in 500ms-long windows wih a 80% overlap, moving he window in 100ms-long seps. To compare he energy of wo signals (he reference one and one using a gain conrol algorihm) we calculaed a squared error-like value by aking he squared sum of he difference beween he energy levels of he corresponding windows: K A, B = i = 1 A () i B () i ref ( energy energy ), diff (11) where energy A(i) is he energy level of signal A in he ih window, and K is he number of windows. One signal will always be he reference recording, hus we ge one value for each recording made. These scores can be easily compared: he lower his number is, he closer he 2 recording is o he reference one, hus he beer i is. As hese values are difficul o read, we calculaed heir relaive error reducion (or RER) scores as well: he appropriae baseline recording had a score of 0%, whereas he reference recording having an energy difference of 0 had a score of 100%. The inermediae values were assigned linearly, e.g. for a baseline energy difference of 5000 and a score of 1500, he RER value will be 70%, as is 70% of 5000, meaning his much of he error was eliminaed. D. Evaluaion via Senence Recogniion Energy levels can be calculaed quickly, and hey can be used very reliably o esimae he difference beween he volume levels of wo recordings. Bu his approach has a serious limiaion: we adjus he gain o make he recording more undersandable; if wo signals have quie differen energy levels, bu boh can be undersood very well, his echnique canno deec i. Unforunaely, undersandabiliy is no a well-defined noion. Hence o assess i, we urned o sandard echniques of speech recogniion. The aim of speech recogniion is o ransform spoken words o wrien ex. In a ypical speech recogniion process, firs feaures [6] are calculaed from he inpu signal, usually on he basis of is specral represenaion [7], which process is called feaure exracion. In he nex sep, following he frame-based approach [8], small, equalsized pars (he so-called frames) are classified independenly and assigned o one of he possible phonemes, which is he phoneme classificaion subask [9]. I is usually done by applying some saisical machine learning algorihm like Gaussian Mixure Models (GMMs) [10] or Arificial Neural Neworks (ANNs) [11]. (The combined seps performed up o his poin are usually called he acousic model.) Nex, based on he resul of he classificaion and he probabiliies of possible wordsequences (which are supplied by a language model) he mos probable word sequence is chosen, which will be he ranscrip of he inpu speech signal. The accuracy value of his process can be deermined by comparing his resul o he real word sequence belonging o he speech signal. As one can see, i is a quie sandard process, which makes i feasible for generaing auomaic measures. One ineresing aspec of his process is ha, similar o he case of human hearing and comprehension, he accuracy value obained decreases when he qualiy of he played recording becomes worse. I is so because in his case he inpu signal conains less and less informaion, which means ha he signal processing and feaure exracion pars calculae more noisy feaures. I produces a less and less reliable phoneme classificaion, finally resuling in word-level misakes. I is known, however, ha curren speech recogniion sysems are much more sensiive o noise han human liseners. 51

6 One migh find i surprising ha his behaviour is acually beneficial for he applicaion of senence-level speech recogniion for measuring he undersandabiliy of a recording, bu his is due o wo main reasons. Firsly, i follows he way human hearing works; and if we wan o measure he amoun of human undersandabiliy, any mehod ha mirrors human hearing is of course helpful. And secondly, i is likely ha, on he receiver side, we do no wan o play he ransmied signal o a human, bu we would like o process i by an auomaic speech recogniion sysem. In his case oday s speech recogniion echnologies are quie capable of measuring he qualiy of he signal played. In pracice we followed he frame-based approach [8]: we divided he speech signal ino small, equal-sized pars, which afer feaure exracion were classified as one of he possible phonemes. We could have measured he qualiy of signals based only on he resul of phoneme classificaion [9]; performing senence recogniion, however, is a higher-level concep, which seems o be more meaningful. We applied he sandard 13 MFCC coefficiens along wih heir derivaives and he second derivaives (MFCC + Δ + ΔΔ for shor) [12] as feaures for phoneme classificaion, and applied Gaussian Mixure Models (GMMs) for i wih 11 componens [10].We performed he raining of hese GMMs on recordings of broadcas news, also recorded by using he wireless sensors, bu his ime using he fixed disance of 50 cenimeers. The feaures were calculaed by uilising he HTK oolki [13]. For language modeling usually a soluion called N-gram modeling [14] is used in speech recogniion sysems. In i he occurrence of all N possible successive words are couned, from which he probabiliy of he Nh word can be deermined; hen he probabiliy of a word sequence can be calculaed by muliplying he probabiliy of each word occurring. This approach unforunaely has he drawback ha a quie big and relevan daabase (in our case wrien exs of broadcas news) is required, hus we oped for anoher, alhough simpler soluion: we simply lised he possible words and allowed any combinaion of hem wih he same probabiliy. Measuring he performance of a coninuous recognizer speech recogniion applicaion is somewha more complicaed han measuring i for an isolaed word recognizer, where word-level accuracy scores can be deermined readily as he number of exac word maches. In he case of senence recogniion he common mehod is o calculae he word-level edi disance of he wo word sequences (he real and he resulan); ha is, we consruc he resuling senence from he real ranscrip by using he following operaions: insering and deleing words, and replacing one word wih anoher one. These operaions have some cos (in our case he common values of 3, 3 and 4 were used, respecively), and hen we pick an operaion se having he lowes cos. Now we can calculae he following measures: and Correcnes Accuracy N S D s = (12) N N S D I =, (13) N where N is he oal number of words in all he original senences, S is he number of subsiuions, D is he number of deleions and I is he number of inserions. Correcness does no ake he number of inserions ino accoun, bu he accuracy value can be negaive, which is why usually boh scores are used. We again calculaed he relaive error reducion scores, where we had several opions of choosing he maximum value for boh he accuracy and correcness values. We could pick 100% as he maximum, as is common in speech recogniion. The drawback of his choice is ha i oally ignores he recording condiions, and assumes ha perfec recogniion can be achieved. The reason for his is ha in he field of speech recogniion usually he bes configuraion is uned, which in our case is he recording we chose for he role of a glass ceiling. In he remaining wo choices we do no consider 100% accuracy as he maximum; insead we ook he performance of a basic recording. Here we could choose eiher he one a he corresponding disance, or he one made a 50cm (having he glass ceiling value). As boh are valid opions, we also calculaed boh raios. The (original) error value is he difference beween he accuracy scores of he basic and he baseline recordings; o express how much of i was eliminaed (which is he RER score), we calculaed he difference beween he accuracy score of he acual and he baseline recordings, and divided i by he error value. These scores are referred o as same disance and glass ceiling RER values, whereas he firs version, described above, was called he absolue RER score. (This process, of course, was repeaed for he correcness values as well.) V. RESULTS Firs we had o se he parameer of boh algorihms o he opimal value: for his we found he inerval of values which worked well by preliminary ess, hen explored i wih a small sep size. The sep parameer of he Equal- Sepping Algorihm (ES) was esed beween 1 and 6, whereas for he Weighed-Averaging Algorihm (WA) w was esed beween 1/32 and 10/32 wih a sep size of 1/32. 52

7 8 6 consan volume (50cm) varying volume (50cm) ES gain conrol (50cm) Energy ime (s) Fig. 1. Energy levels of he reference recording (grey coninuous line), he baseline recording (grey rugged line) and he ES algorihm wih sep = 3. A. Resuls Using he Energy Level We found ha he bes parameer values were sep = 3 and w = 5/32 when evaluaing he recordings in erms of heir energy level. Firs he full energy raio was calculaed for each recording (see Table 3). As can be seen, he disance clearly affecs he energy raios when we do no use a gain conrol: he recordings made a 20cm had a 50% bigger oal energy han he reference recording, whereas he recordings made a 100 and 200 cenimeers had significanly less. Boh gain conrol algorihms, however, could indeed compensae for he overall loudness (or quieness) a hese disances, resuling in a oal energy raio very close o 1. For he whole parameer inervals esed, his raio varied beween 0.77 and 1.08 for he ES algorihm and beween 0.87 and 1.11 for he WA mehod. TABLE 3. Full energy raios of he wo gain conrol algorihms (ES and WA). Disance Recording 20cm 50cm 100cm 200cm Basic Baseline ES AGC, sep = WA AGC, w = 5/ The energy level difference diff values, calculaed according o (11), can be seen in Table 4, while he relaive error reducion scores are shown in Table 5. I is no surprising ha he diff values of he basic and baseline recordings increase when he disance changes from he opimal one. The only excepion is he baseline recording a 100cm: i has a lower diff score han a 50cm, which is probably due o he high number of loud pars in he signal played. I may also be why a smaller score is obained for he baseline signal han for he basic one a 200cm, leading o he negaive RER score for he laer. The basic recording from 50cm has an excepionally small difference value due o he indeerminisic naure of recording (i.e. i was no exacly he same as he reference one). Boh gain conrol mehods, however, performed quie well. As he disance varied from he opimal one, he diff values increased slighly, bu he RER scores reflec he fac ha using gain conrol was an effecive way of counering his effec. The % and % RER values (ES and WA algorihms, respecively) are quie good, and in almos every case hese are higher scores han hose of he basic recordings. The only excepion is a 50cm, bu i was pracically impossible o bea his score (99.93%) here, and he values exceeding 70% are also quie saisfacory. TABLE 4. Energy differences of he wo basic recording ypes and of he wo gain conrol algorihms (ES and WA) relaive o he reference recording. Disance Recording 20cm 50cm 100cm 200cm Basic Baseline ES AGC, sep = WA AGC, w = 5/ TABLE 5. Energy difference relaive error reducions scores of he wo basic recording ypes and of he wo gain conrol algorihms (ES and WA). Disance Recording 20cm 50cm 100cm 200cm Basic 47.39% 99.93% 55.82% -9.89% Baseline 0.00% 0.00% 0.00% 0.00% ES AGC, sep = % 73.85% 66.44% 66.87% WA AGC, w = 5/ % 70.71% 62.12% 69.82% Visually inspecing he energy levels a 50cm using he ES algorihm wih sep = 3 (see Figure 1), we may say ha he algorihm seems o be quie effecive. (The WA mehod produced a very similar curve wih w = 5/32.) While he energy levels of he baseline recording grealy differ from he reference one, he gain conrol algorihm compensaed for he jumps in volume: i usually differs from he reference recording by only a small amoun. The only weakness of he mehod seems o be he periods afer longer silences, where i resuled in much higher energy values han hose of he reference. Figure 2. shows he energy levels of he basic recordings (in he upper box), and of he ES algorihm wih sep = 3 (in he lower box) a each disance esed. I can be clearly seen ha he disance beween he sensor and he sound source srongly affecs he energy levels when here is no gain conrol: he four corresponding curves are quie differen from each oher. (Noe ha energy is displayed on a log-scale.) On he oher hand, he energy levels of he recordings using a gain conrol algorihm fall fairly close o 53

8 8 6 20cm 50cm 100cm 200cm Energy ime (s) 8 20cm 50cm 100cm 200cm 6 Energy ime (s) Fig. 2. Energy levels of he basic recording (up), and he signal wih varying volume using he ES algorihm wih sep = 3 (down) a differen disances. each oher, indicaing ha he mehod was able o amplify sources having differen volumes o roughly he same level, which agrees wih our previous findings involving oal energy raios. B. Resuls Using Senence Recogniion Unlike evaluaing he gain conrol mehods in erms of he energy levels, we found no definie bes parameer values when performing senence recogniion. Overall, we chose he parameers sep = 2 and w = 7/32, bu we shall discuss his issue more hroughly laer. The resuling correcness values can be seen in Table 6, while he corresponding accuracy scores are given in Table 7. We achieved 72.59% and 70.17% on he reference recording for correcness and accuracy, respecively, which are indeed very close o hose of he basic recording made a 50cm (which were is as idenical as was pracically possible o he reference one). Considering he very noisy recordings due o he small microphone on he sensor board, and he simpliciy of he language model (which usually significanly aids he speech recogniion process [15]), we found his score surprisingly high. TABLE 6. Correcness resuls of he wo gain conrol algorihms (ES and WA). The reference recording produced a score of 72.59%. Disance Recording 20cm 50cm 100cm 200cm Basic 67.07% 72.24% 52.41% 5.52% Baseline 54.48% 59.31% 38.28% 7.07% ES AGC, bes values 74.31% 68.45% 43.45% 10.00% ES AGC, sep = % 65.86% 41.21% 10.00% RER (same disance) % 50.66% 20.74% RER (glass ceiling) % 50.66% 8.63% 4.50% RER (absolue) 43.56% 16.10% 4.75% 3.15% WA AGC, bes values 75.17% 68.79% 45.34% 10.17% WA AGC, w = 7/ % 64.48% 44.48% 10.17% RER (same disance) % 39.98% 43.88% RER (glass ceiling) % 39.98% 18.26% 4.76% RER (absolue) 45.45% 12.71% 10.05% 3.34% TABLE 7. Accuracy resuls of he wo gain conrol algorihms (ES and WA). The reference recording produced a score of 70.17%. Disance Recording 20cm 50cm 100cm 200cm Basic 63.28% 69.48% 50.17% 5.34% Baseline 50.17% 56.72% 35.69% 6.55% ES AGC, bes values 70.69% 65.34% 40.86% 8.45% ES AGC, sep = % 63.10% 37.93% 8.45% RER (same disance) % 50.00% 15.47% RER (glass ceiling) % 50.00% 6.63% 3.02% RER (absolue) 40.84% 14.74% 3.48% 2.03% WA AGC, bes values 72.41% 65.52% 43.10% 9.14% WA AGC, w = 7/ % 64.48% 42.76% 9.14% RER (same disance) % 60.82% 48.83% RER (glass ceiling) % 60.82% 20.94% 4.12% RER (absolue) 44.63% 17.93% 10.99% 2.77% The performance of he basic and baseline recordings clearly show ha, among he four esed cases, he disance of 50 cm could be considered as he for boh cases: he accuracy and correcness scores are he highes using his disance, while hey fall when he sensor is closer or furher away from he microphone. The exremely small scores of he recordings made a 200cm, however, are surprisingly low. Our previous ess involved phoneme recogniion in he same circumsances as we had here [16], and hey also showed a decrease in he phoneme classificaion scores a his disance, bu by a much smaller amoun (hey fell from he 83.19% glass ceiling level o 53.75% and 53.05%, basic and baseline recordings, respecively). I could be, however, ha his decrease in he phoneme idenificaion performance made he acousic model of speech recogniion unreliable in pracice, which brough he senence-level recogniion performance down o his level. This hypohesis is also corroboraed by he small difference in he phoneic accuracy of he wo recordings, suggesing ha due o he low volume (caused by he very large disance), mos pars of he uerances were jus no disinguishable from background noise (i.e. silence). Low speaker volume could also be he reason for he higher speech recogniion accuracy for he baseline 54

9 recording han ha of he basic one: he baseline recording had a varying volume, so in a number of cases i was louder, aiding he undersanding process in his siuaion, whereas he lower-volume pars did no degrade accuracy any furher. (Noe ha his resul mirrors our findings when we evaluaed hese wo recordings via he use of energy levels: he baseline recording also performed beer han he basic one measured by ha evaluaion meric.) I also means ha calculaing he relaive error reducion using he basic score a he same disance makes no sense in his case. Examining he performance of he gain conrol algorihms, perhaps he mos ineresing finding is ha, unlike he evaluaion done via energy levels, we could no find an opimal parameer value for eiher algorihm. Usually he seings which performed bes a 50 and 100 cenimeers produced worse senence recogniion scores a 20 and 200 cenimeers han he oher cases, which, in conras, worked subopimally when using he former disances. The reason for his is probably ha a 50 and 100 cenimeers he volume has a relaively small variaion, which prefers mehods wih smaller, smooher changes. On he oher hand, recordings made a 20 and 200 cenimeers require more flexible gain conrol mehods, which allow greaer jumps in he gain. Of course, for a recording applicaion i is unreasonable o expec consan swiching beween parameer values while recording, hus we chose one parameer seing for boh algorihms ha we considered bes; bu we also lised he bes scores achieved for boh algorihms and for all four disances in all he parameer inervals esed. Looking a he scores, we may conclude ha we could achieve significan improvemens by using eiher of he gain conrol algorihms described here. The only excepion was when we made recordings a 200 cenimeers, where, despie he relaively high error reducion scores, he performance scores sill remain a he unusable level. I is probably because raining was performed on recordings made from a fixed disance (50 cenimeers). In heory gain conrol could couner his effec by raising microphone sensiiviy, bu he Signal-o-Noise Raio (SNR) [17] canno be raised his way, because he background noise is also amplified. In he oher cases, however, he relaive error reducion scores based on using he same disance are quie convincing (ranging from 20.74% o % and 15.47% o %, correcness and accuracy, respecively), and he oher wo RER values are also good. A noeworhy case is using gain conrol when recording a 20 cenimeers, where even he glass ceiling value could be exceeded. These resuls imply ha wha we regarded as an recording environmen (using a consan volume-level recording from he bes disance) was no he bes one possible. The reason for his is probably ha in human speech, even wihou arificially inroducing volume level changes (as we did when consrucing he baseline recordings and he ones recorded using gain conrol), here are also volume level changes presen [18] ha could also be handled by he use of gain conrol. Overall, alhough here are small differences in he performance of he wo mehods, boh achieved similarly good resuls, hence boh can be recommended for pracical use. The small advanage of he Weighed-Averaging Gain Conrol Algorihm over he Equal-Sepping one could be due o is smooher ransiion of gain level and is abiliy o make bigger jumps when required. Also, i was demonsraed ha gain conrol could indeed be effecive when using wireless sensors o record speech daa, since he undersandabiliy of he sound signals ransferred (measured in erms of speech recogniion senence-level error scores) improved significanly. VI. CONCLUSIONS Wireless sensors are recen, low-capaciy devices used for monioring heir immediae environmen, which includes recording and ransmiing audio informaion. In his siuaion, however, here could be a big difference in he volume level of he observed signals due o he presence of muliple and/or moving speakers. Varying volume can indeed harm he undersandabiliy of signals, which can be compensaed for by applying Auomaic Gain Conrol mehods. The wo algorihms inroduced in his work (which were designed o mee he paricular requiremens of our se-up, bu could also be used elsewhere) proved successful when we measured heir performance using he difference in volume levels and when applying senencelevel auomaic speech recogniion. Overall, he qualiy of a sound recording made could be improved significanly using some AGC echnique like our soluions. ACKNOWLEDGMENTS This sudy was parially suppored by he TÁMOP /08/1/ programme of he Hungarian Naional Developmen Agency. REFERENCES [1] S. Ramanahan and M. Seensrup, A survey of rouing echniques for mobile communicaions neworks, Mobile Neworks and Applicaions, vol. 1, no. 2, pp , [2] S. Kang, H. Choi, H. Yoon and K. Park, Auomaic gain conrol for he uniform ampliude of inerferen signal in a Laser Doppler Vibromeer, Proceedings of SICE-ICASE Inernaional Join Conference SICE-ICCAS 2006, Busan, Souh Korea, pp , [3] C. S. Bae, An auomaic pacemaker sensing algorihm using auomaic gain conrol, Proceedings of he 1999 IEEE Region 10 Conference TENCON 99, Cheju Island, Souh Korea, pp , [4] Z. Tüske, P. Mihajlik, Z. Tobler and T. Fegyó, Robus Voice Aciviy Deecion based on he enropy of noise-suppressed specrum, 9h European Conference on Speech Communicaion and Technology Inerspeech Eurospeech, Lisboa, Porugal, pp ,

10 [5] J.M. Górriz, J. Ramírez, J.C. Segura and S. Hornillo, Voice Aciviy Deecion using higher order saisics, Proceedings of he 8 h Inernaional Work-Conference on Arificial Neural Neworks IWANN 2005, Barcelona, Spain, pp , [6] L. R. Rabiner and R. W. Schafer, Digial Processing of Speech Signals, Prenice-Hall, New Jersey, [7] J. Benesy, M. M. Sondhi and Y. Huang, Springer Handbook of Speech Processing, Springer-Verlag, Berlin, [8] L. Rabiner and B.-H. Juang, Fundamenals of Speech Recogniion, Prenice-Hall, New Jersey, [9] A. Kocsor, L. Tóh, A. Kuba Jr., K. Kovács, M. Jelasiy, T. Gyimóhy and J. Csirik, A comparaive sudy of several feaure space ransformaion and learning mehods for phoneme classificaion, Inernaional Journal of Speech Technology, vol. 3, no. 3/4, pp , [10] R. O. Duda and P. E. Har, Paern Classificaion and Scene Analysis, Wiley & Sons, New York, [11] C. Bishop, Neural Neworks for Paern Recogniion, Clarendon Press, Oxford, [12] X. Huang, A. Acero and H.-W. Hon, Spoken Language Processing, Prenice Hall, New Jersey, [13] S. Young, The HMM Toolki (HTK) (sofware and manual), hp://hk.eng.cam.ac.uk/, [14] F. J. Damerau, Markov Models and Linguisic Theory: An Experimenal Sudy of a Model for English, Mouon De Gruyer, Berlin, [15] F. Jelinek, Saisical Mehods for Speech Recogniion, The MIT Press, Cambridge, USA, [16] G. Goszolya, D. Paczolay and L. Tóh, Auomaic Gain Conrol algorihms for wireless sensors, Proceedings of he Inernaional Join Conference on Compuaional Cyberneics and Technical Informaics ICCC-CONTI 2010, Timisoara, Romania, pp , [17] H. G. Hirsch and C. Ehrlicher, Noise esimaion echniques for robus speech recogniion, Proceedings of he Inernaional Conference on Acousic, Speech, and Signal Processing ICASSP 1995, Deroi, USA, pp , [18] M. Holmberg and D. Gelbar, Auomaic Speech Recogniion wih an adapaion model moivaed by audiory processing, IEEE Transacions on Audio, Speech and Language Processing, vol. 14, no. 1, pp , Manuscrip received Sepember 8, 2010; revised December 4, 2010; acceped for publicaion April 7,