Improvement of Microphone Array Characteristics for Speech Capturing

Transcription

1 odern Applied cience; Vol. 9 o. 6; 05 I E-I Published by Canadian Center o cience and Education Iproveent o icrophone Array Characteristics or peech Capturing tolbov ikhail Borisovich & Aleinik ergei Vladiirovich t. Petersburg ational Research University o Inoration Technologies echanics and Optics Russia Correspondence: tolbov ikhail Borisovich t. Petersburg ational Research University o Inoration Technologies echanics & Optics 970 t.petersburg Kronverkskiy pr 49 peech Technology Center Russia t.petersburg Krasutskogo street 4 Russia. E-ail: stolbov@speechpro.co/aleinik@speechpro.co Received: eceber 8 04 Accepted: January 5 05 Online Published: ay doi:0.5539/as.v9n6p30 URL: Abstract This paper presents a new adaptive technique or speech capturing in adverse conditions using icrophone arrays. The proposed technique is based on requency-doain alignent o icrophone signals with the output o the ixed beaorer directed to the target speaker. This alignent procedure iproves pattern directivity and reduces sidelobes. The low coplexity o the technique is achieved by eans o a requency-doain ipleentation o the algorith. This akes it possible to ipleent this technique in real-tie applications with a large nuber o icrophones. The technique was evaluated on speech data corrupted by varying levels and directions o noise and intererence. The proposed technique iproves the directivity pattern o a traditional ealy & u beaorer as well as provides additional suppression o spatially incoherent noise diuse noise and intererence with inial loss o the target signal quality. Keywords: icrophone array speech processing noise reduction adaptive beaoring. Introduction icrophone arrays As are an eective tool or any applications speech transcription voice/speaker identiication etc.. As also enable long-distance speech capture in adverse conditions Zang et al. 00. For speech enhanceent the ain advantage o using icrophone arrays rather than a single icrophone is that a icrophone-array-based beaorer can spatially suppress ultiple interering signals and dierent background noises diuse noises independent noises etc. while aintaining iniu distortion o the target signal ro the look direction Gannot et al.00. A great nuber o well-known ixed or adaptive algoriths or signal processing in A are described in Widrow & tearns 985 and Brandstein & Ward 00. Coparison o dierent beaoring techniques ay be ound or exaple in VanVeen & Buckley 988. ierent approaches can be classiied into three ain categories Fischer & ier 995: Conventional beaoring Kellerann 99 ahieux et al Adaptive beaoring Frost 97 oshuyaa et al.999. Adaptive postiltering Zelinski 988 ier et al.00. A conventional ixed beaorer FBF uses known elay & u or Filter & u algoriths Flanagan et al. 985 enables the suppression o both coherent intererence and statistically independent noise. owever not always with enough good eiciency. ierent adaptive algoriths Frost 97 Griths & Ji 98 etc. can be used or ore eicient coherent intererence suppression while adaptive postiltering Löllann & Vary 007 Cao et al. 994 can be applied or incoherent noise cancellation. everal dierent integrated techniques were developed or the suppression o both coherent and incoherent noise see or exaple Ortega-García & González-Rodríguez 996. owever practical application o these algoriths presents a nuber o diiculties: audible usic noise target signal suppression sensitivity to icrophone isatch oclo & oonen 007 etc. In our previous paper tolbov & Aleinik 04 we proposed a new adaptive bea-oring algorith. It was developed or an 8-eleent A and its eiciency was proved in practice by experiental studies on artiicial odel and real signals. This paper presents a odiication o that algorith as well as a detailed study o the algorith and its 30

2 odern Applied cience Vol. 9 o. 6; 05 characteristics advantages and disadvantages.. The Proposed Technique ere we briely describe the odiication o our ethod proposed in tolbov & Aleinik 04.. The escription o the odiied ethod The ethod is based on requency doain alignent o icrophone signals with the output o a ixed beaorer directed to the target speaker. The block diagra o the ethod exaple or a 4-icrophone array is shown in Figure. We ipleent this procedure in rae-based requency doain ier & Wasilje 99. The three initial steps are well-known as a requency-doain elay & u FBF. At the irst step the input signals o each icrophone x t are transored into spectra X k using short-tie Fourier TFT k ' k x 0 t TFT delay 5 ilters V 0 0 k 6 x t x t TFT TFT V V k k Σ IFT yt x 3 t TFT V 3 3 k 4 k calculation 3 Σ s k Figure. Block diagra o the proposed technique. where is the icrophone index is the requency bin index and k is the rae index Figure block. Then each signal X k is delayed in block by ultiplication with the coplex steering vector V : k V X k where is the desired source direction. At the third step the requency-doain FBF output is calculated block 3 using the ollowing equation: s k k. We point here that the signals and represent an ideal situation where icrophone characteristics are equal to each other. In practice there are luctuations in icrophone sensitivity and phase both are generally requency dependent. To copensate or these luctuations we irst estiate the transer unction k : Φˆ k Φˆ s k 3 k where Φ ˆ k and ˆ k are the estiated cross spectra and the power spectra o the s Φ k and s k signals. These estiates are obtained using exponential rae-by-rae averaging: ˆ ˆ Φ k α Φ k α k k s s s ˆ ˆ Φ k α Φ k α k k 4 3

3 odern Applied cience Vol. 9 o. 6; 05 where α T τ is the tie constant o the decaying T and τ are the rae duration and adaptation tie constant respectively. The calculation o 3-4 is perored in block 4 Figure. Then the atched transer unction k is transerred to block 5 where the icrophone signals are transored as: ' k k k. 5 Ater odiication the new output spectru is calculated in block 6 as ollows: ' out k. k 6. Analysis o the Proposed Algorith agnitude and Phase Inluence We should point out here that urther on we oit the rae index k or the sake o siplicity. Consider again the A directed perpendicularly to the line o icrophones. aturally the transer unction 3 is a coplex unction i.e. we can write: exp j arg 7 where and arg are the agnitude and the phase o the unction respectively. It is clear that generally both coponents agnitude and phase aect the results 5 and 6. Transoring in 5 to only the phase o or only the agnitude o allows us to investigate these inluences separately. Results these studies are presented below... Phase Inluence The physical eaning o the inluence o the phase can be easily explained as ollows. Consider a 3-eleent equidistant linear A. In that case the iddle second icrophone represents the phase center o the A. I our A is directed perpendicularly to the line o icrophones then every eleent o the steering vector is real valued and equal to. Accordingly the A output s is a siple su o three icrophone output signals X 3 divided by 3. ow let the input signal be a haronical plane wave ro soe direction. It is clear that in this case phase dierences between s and X are equal to ϕ 0 -ϕ or the irst second and third icrophone respectively. oreover exactly the sae phase dierences are obtained when we calculate cross-spectra 4. o or our 3-eleent A the phase o can be written as: arg ϕ; arg 0; arg 3 -ϕ. Thus when we calculate 5 we provide additional rotation o the X and X3 signals to angles ϕ and -ϕ i.e. phase dierences are doubling. As a result or the input signal requency we have the A directivity pattern corresponding to the requency : the width o the ainlobe and sidelobes are two ties narrower the sidelobes are saller etc. O course this conclusion is valid or any nuber o icrophones. Figure shows the resulting directivity patterns o an 8-icrophone A or a Conventional elay & u FBF and the proposed ethod when only phase was used in 5. The distance between the icrophones was 5 c the input signal was a 000 z haronical plane wave. Figure. irectivity patterns or elay & u and the proposed ethod or arg 3

4 odern Applied cience Vol. 9 o. 6; 05 It is clear that the dashed curve corresponds to the input signal with the requency equal to 4000 z while in reality the input signal requency was 000 z... agnitude Inluence I we use only the agnitude o in 5 the result will be qiute dierent. I the input signal is a plane wave ro an angle then the aplitude Aout o the output signal out can be written as: Aout Ain 8 where A in is the input signal aplitude and is the agnitude directivity pattern o the A. A close inspection o 5 and 6 shows that in this case E 9 where E denotes the atheatical expectation. Rewriting 8 and 9 in ters o expectations we get: E Aout E Ain E Ain 0 This eans that the resulting directivity pattern is equal to the square o the initial one. These conclusions are conired by Figure 3 which shows the resulting directivity patterns or the sae paraeters o the icrophone array and the input signal as in Figure when only the agnitude o is used in 5. Figure 3. irectivity patterns or elay & u and the proposed ethod or. We can see that ain lobe o the dashed curve is slightly narrower and the level o sidelobes or the dashed curve on the db scale is two ties lower than that or the solid curve. I we copare Figs. and 3 we can conclude that the phase ainly reduces the width o the ain lobe while the agnitude has a greater eect on the aplitude o the sidelobes. O course ainlobe width reduction looks very useul. oreover our investigation have shown that the use o the coplex allows to get both ainlobe width and sidelobes reduction. owever the data in Figs. and 3 were obtained using haronic signals without interering noise i.e. in ideal conditions. On the other hand our experients showed that irst the phase ethod works poorly in the presence o high level additive noise. econd the use o the phase increases parasitic usical noise in the output speech signal. In the end we let the phase ethod or urther research and ocused on the investigation o 3 with agnitude in the nuerator i.e. everywhere below we assue that or our proposed ethod: Φˆ k Φˆ 33 k k.3 Analysis o the Proposed Algorith or Coherent oise Consider coherent noise suppression with the presence o a target signal. The icrophone output signal in this case equals intererence plus speech signal: s

5 odern Applied cience Vol. 9 o. 6; where is the output icrophone signal; is target signal coing ro the look direction; is intererence coing ro the direction. Conventional FBF output can be written as:. 3 In this case the transer unction can be written as:. 4 where is the tie averaging sybol. ence we get: 5 where: is the ignal-to-intererence ratio. I the target speech signal is stronger than the intererence we get which eans that the target signal and the noise coe ro the FBF output unchanged and the relation reains the sae copared with FBF. In case there is no target signal and according to 5 we can draw the sae conclusion as in 0 i.e. noise reduction is equal to: 6 Thereore the proposed ethod doubles noise suppression in the db scale. For an 8-eleent A with kz intererence the level o the irst sidelobe is equal to -3 db. Ater the application o the proposed algorith the level is reduced to -6 db..4 Analysis o the Proposed Algorith or Incoherent oise For independent icrophone noise we get: 7 where is the noise on the -th icrophone such that. In this case the FBF output can be expressed as: 8 In this case: and hence:. 9 Thus or dierent we get: i ; i. Thereore in the presence o a speech signal the siple FBF signal appears at the output. On the other hand white noise in speech pauses i.e. when is additionally suppressed by the actor o. I М8 suppression aounts to 9 db with total suppression o 8 db. s s IR IR IR > >> IR IR 0 IR > R R k E k 0 s s R R s R > >> R R / > << << R

6 odern Applied cience Vol. 9 o. 6; Experients and Results 3. Equipent and Paraeters In our experients we investigated the proposed ethod using 8 equally spaced icrophone array with 5 c inter-icrophone spacing and 35 c total aperture length. Oni-directional icrophones Knowles Electronics were used. The signals o the icrophones were sapled with the requency F s 6 kz. We used a standard Overlap-and-Add OLA technique with the rae length o 5 saples 50% overlapping and a ann window. The tie constant α in 4 was about In practice one o the detected issues was audible usic noise caused by estiated spectru rando luctuations 4. It is known that it is ostly low-requency estiated spectru luctuations that cause this type o noise. To reduce the usic noise we propose the ollowing transer unction liitations: i > ax then ax 0 i < in then in We used ixed ax ax is set by user. Estiated directivity patterns as well as values o noise reduction R and intererence reduction IR were used or evaluating the perorance o the proposed ethod. White Gaussian noise WG and Wide-Band Gaussian noise WBG with a z band were used as input signals. We investigated the ethod or three scenarios: incoherent coherent and diuse noises. Testing was conducted with both odel signals and real signals recorded in an oice. For coparison we also estiated the characteristics o a conventional FBF. 3. Results 3.. irectivity Patterns irectivity patterns Ps shown in Figures and 3 were obtained using an artiicial single-requency haronic signal and thereore do not provide coplete inoration about the perorance o the proposed ethod. For clariication we conducted a series o experients using the odel and real WBG as input signal or A. Figure 4 shows the Ps when the input signal is artiicial WBG ro the direction 0. It can be seen that or our WBG oscillating sidelobes are absent and the Ps gradually decrease as the A angle oves away ro the signal direction 0. We can also see that the width o the ainlobe is large enough. This act can be explained by the eect o low-requency coponents o WBG as our antenna has a length o 35 c which provides poor directivity at requencies below 000 z. Using our ethod we obtain a siilar behavior o the P curve dashed curve but the ainlobe is narrower and the sidelobes are saller. Figure 4. irectivity patterns or elay & u and the proposed ethod or artiicial wide-band Gaussian noise bandwidth is z 35

7 odern Applied cience Vol. 9 o. 6; 05 Figure 5 shows the Ps or the signal obtained in a real experient: WBG noise eitted by high-quality speakers located at a distance o 4 eters ro the A in an oice with the reverberation constant equal to 860 s. Figure 5. irectivity patterns or elay & u and the proposed ethod or wide-band Gaussian noise recorded in a real oice We can see that despite the act that the overall A perorance deteriorates the proposed ethod has better characteristics than the conventional elay & u FBF. 3.. oise Reduction as a Function o Frequency We already entioned the well-known act that noise reduction in a A depends on noise requency. Consider Figure 6 which shows three power spectra. In this experient WG eitted by high-quality speakers located 6 ro the A and hidden ro the A by a screen siulated diuse noise in the oice. Figure 6. Power spectra o input and output signals o the A or diuse noise The solid curve is the power spectru o the single icrophone signal; the dotted and dashed curves are the power spectra o the A output signal when elay & u and the proposed ethods o BF are used 36

8 odern Applied cience Vol. 9 o. 6; 05 respectively. The igure shows that all three curves are alost equal to each other in the low-requency doain z. Then the dotted and the dashed curves begin to go down indicating that both elay & u and the proposed algoriths start to suppress the diuse noise. The lower the curves are in coparison with the solid one the stronger the noise suppression. o Figure 6 clearly deonstrates the advantages o the proposed ethod copared to the elay & u FBF. A slight convergence o the curves in the region above 6000 z can be explained in our opinion by the appearance o parasitic sidelobes with a high aplitude. We should point out that the data shown in Figure 6 are very useul as they allow us to calculate the level o diuse noise reduction in the requency bands. For exaple let us denote the power spectru o the single icrophone signal solid curve as P in and the power spectru o the A output signal or the proposed ethod as P out. Then noise reduction in the discrete requency band can be written as: R Δ 0 log 0 0 P 0 log in 0 P out 3..3 oise Reduction or iuse oise and or Intererence We calculated noise reduction or three types o noises: odel incoherent noise real diuse noise and real intererence. odel incoherent noise is a set o 8 independent artiicial rando sequences with Gaussian distribution generated using the WELL04a algorith described by Panneton et al iuse noise is a real WG recorded by the A in an oice and generated as explained in the description o Figure 6. Intererence is also a real recorded WG in an oice coing to the A ro the angle o 40 degrees when the A is oriented at 0. Estiations o R and IR were provided in the requency band z. Table. Estiated noise and intererence reduction oise type FBF Proposed ethod Incoherent noise 9 db 8 db Coherent noise 40 0 db 7 db iuse noise 6.9 db.9 db The results presented in Table show that the proposed ethod has considerable advantages copared to FBF. 4. iscussion In this paper we proposed a novel requency-doain alignent technique o speech enhanceent or directional icrophone array systes. The proposed technique iproves the A directivity pattern copared to the conventional elay & u FBF: we show both analytically and experientally that the proposed ethod narrows the ainlobe and reduces the sidelobes o the P. Two odiications o the proposed ethod phase and agnitude odiications were considered and studied in detail. We perored a nuber o experients using real eitted and recorded in a real oice and siulated acoustic signals. Various types o signals and noises were used in the experients: haronic signals white Gaussian noise wide-band Gaussian noise z etc. We achieved a good siilarity o theoretical and experiental results. Our experients showed that or an 8-icrophone array it is possible to achieve noise reduction o up to.9 db or diuse noise and up to 8 db or incoherent noise. 5. Conclusion The proposed technique showed good results in the processing o acoustic data. oreover the technique which involves a low-coplexity algorith can be used in systes with a large nuber o icrophones or exaple a two-diensional icrophone array in real-tie applications. We think that the technique can be cobined with several types o icrophone array processing algoriths e.g. null-steering post-iltering etc. In a liited nuber o experients we ound that the cobined use o the null-steering algorith and the proposed ethod enhances the suppression in the null direction. We also ound that the syste consisting o our algorith and a sequentially connected Zelinsk post-ilter perors better than either o these coponents separately. owever we did not conduct detailed studies in this area. We should point out that usical noise is one o the ajor probles liiting the perorance o the proposed ethod. Cobination with other adaptive requency-doain algoriths as well as usical noise suppression and detailed study o the perorance o the phase and coplex transer unction are a task or uture work. Acknowledgents The work was inancially supported by the inistry o Education and cience o the Russian Federation 37

9 odern Applied cience Vol. 9 o. 6; 05 Contract I RFEFI5754X0033 and by the Governent o the Russian Federation Grant 074-U0. Reerences Cao. & ridharan oody Post-icrophone-Array peech Enhanceent with Adaptive Filters or Forensic Applications. Proceedings. o International yposiu on peech Iage Processing and eural etworks pp oclo. & oonen uperdirective Beaoring Robust Against icrophone isatch. IEEE Trans. on Audio peech and ignal Processing Brandstein. & Ward. B. 00. icrophone Arrays: ignal Processing Techniques and Applications. pringer-verlag. Fischer. & ier K. U An adaptive icrophone Array or ands-free Counication. Proceedings o International Workshop on Acoustic ignal Enhanceent pp. -4. Flanagan J. L. Johnston J.. Zahn R. & Elko G. W Coputer steered icrophone arrays or sound transduction in large roos. J. Acoust. oc. Aer Frost O. L. 97. An algorith or linearly constrained adaptive array processing. Proceedings o IEEE Gannot. Burshtein. & Weinstein E. 00. ignal enhanceent using beaoring and nonstationarity with applications to speech. IEEE Trans. ignal Processing Griths L. J. & Ji C. W. 98. Alternative approach to linearly constrained adaptive beaoring. IEEE Transaction on Antennas Propagation AP oshuyaa O. ugiyaa A. & irano A A robust adaptive beaorer or icrophone arrays with a blocking atrix using constrained adaptive ilters. IEEE Trans. ignal Processing Kellerann W. 99. A sel steering digital icrophone array. Proceeding o the Internat. Conerence on Acoustics peech and ignal Processing pp Löllann. W. & Vary P Post-ilter design or superdirective beaorers with closely spaced icrophones. Proceedings o IEEE Workshop on Applications o ignal Processing to Audio and Acoustics pp ahieux. Le Tourneur G. Gilloire A. aliou A. & Jullien J. P A icrophone array or ultiedia workstations. Proceeding. o the 3rd InternationalWorkshop on Acoustic Echo Control pp Ortega-García J. & González-Rodríguez J Overview o speech enhanceent techniques or autoatic speaker recognition. Proceedings o International Conerence on poken Language Processing pp Panneton F. L Ecuyer P. & atsuoto Iproved Long-Period Generators Based on Linear Reccurences odulo. AC Transactions on atheatical otware 3-6. ier K. U. Bitzer J. & arro C. 00. Post-iltering techniques. In icrophone arrays. Brandstein and.ward Eds. pringer-verlag Berlin eidelberg ew ork ch. 3 pp ier K. U. & Wasilje A. 99. Adaptive icrophone arrays or noise suppression in the requency doain. Proceedings o econd Cost 9 Workshop on Adaptive Algoriths in Counications pp tolbov. & Aleinik. 04. peech Enhanceent with icrophone Array Using Frequency-oain Alignent Technique. Proceedings o 54th Conerence on Audio Forensics Techniques Technologies and Practice pp. -5. VanVeen B.. & Buckley K Beaoring: A versatile approach to spatial iltering. IEEE Acoust. peech ignal Processing pp Widrow B. & tearns Adaptive ignal Processing. ew Jersey: Prentice-all Inc. Zang C. Tao. & ansen J.. L. 00. icrophone array processing or distance speech capture: a probe study on whisper speech detection. Proceedings o Asiloar Conerence on ignals ystes & Coputers pp Zelinski R A icrophone array with adaptive postiltering or noise reduction in reverberant roos. 38

10 odern Applied cience Vol. 9 o. 6; 05 Proceeding o the Internat. Conerence on Acoustics peech and ignal Processing pp Copyrights Copyright or this article is retained by the authors with irst publication rights granted to the journal. This is an open-access article distributed under the ters and conditions o the Creative Coons Attribution license 39