Enhancement of noisy speech signal based on variance and modified gain function with PDE preprocessing technique for digital hearing aid

Transcription

1 33 Journal o Scieniic & Indusrial Research J SCI IND RES VO 70 MAY 0 Vol. 70, May 0, pp Enhancemen o noisy speech signal based on variance and modiied gain uncion wih PDE preprocessing echnique or digial hearing aid D Deepa* and A Shanmugam Deparmen o ECE, Bannari Amman Insiue o Technology, Sahyamagalam , India Received 04 January 0; revised March 0; acceped 8 March 0 This paper deals wih a single-channel speech enhancemen algorihm, preprocessed using parial dierenial equaion, used o overcome degradaion o noisy speech signals in digial hearing aids. Adapive hreshold is esimaed using variance in ime index. Gain is modiied based on adapive hreshold esimaed in requency bins and used o enhance noisy signal. By proposed mehod, deinie improvemen in and reduced MSE can be obained compared o convenional mehod. This mehod provides a greaer degree o lexibiliy and conrol on noise subracion levels ha reduce ariacs in enhanced speech, resuling in improved speech qualiy and inelligibiliy. eywords: MSE, PDE,, Adapive hreshold, Gain uncion, Variance Inroducion To undersand speech wih background noise is diicul even or normal hearing people; his is especially rue or a large number o elderly people and or sensorineural impaired persons. People wih sensorineural loss may need a 5-5 db higher signal-o-noise raio () han normal hearing persons. Hearing impaired paiens cases are as ollows: due o conducive losses, 0%; due o sensorineural losses, > 50%; and cases o mixed origin, 30%. In single channel sysem, speech enhancemen is a challenging on because reerence noise signal will no be available or enhancemen. Convenional power specral subracion mehods -5 or single channel speech enhancemen subsanially reduces noise levels in noisy speech bu i inroduces musical noise. In some cases, improvemen in qualiy migh be accompanied by a decrease in inelligibiliy due o he disorion impared o clean speech resuling rom excessive suppression o acousic noise. In proposed mehod, ollowing preprocessing, which is done by using parial dierenial equaion (PDE), algorihm is based on raio o variance o noisy specrum o is average. *Auhor or correspondence deepa_dhanaskodi@yahoo.co.in Experimenal Secion Parial Dierenial Equaion (PDE) Technique Firs sep in speech enhancemen using PDE is o obain gradien (g) o each sample in noisy speech signal using samples beore and aer curren sample as g g b S( x x, ) S( x, ) S( x + x, ) S( x, ) () where, S(x,) is noisy speech signa x is sampling rae. Aer gradien is calculaed, inluencing coeiciens (ICs) in each direcions o curren sample are compued as IC IC b g + k g b + k () where, IC is orward ICand IC b is backward IC. In Eq. (), k is consan value beween and 00. From ICs and gradiens, speech signal is enhanced as S x, + ) S( x, ) + ( g IC + g IC ) (3) ( b b

2 DEEPA & SHANMUGAM: ENHANCEMENT OF NOISY SPEECH SIGNA FOR DIGITA HEARING AID 333 Fig. Proposed noise reducion algorihm where S(x,) is inpu noisy speech signa is a coeicien beween 0. o 0.6 represening sep o noise reducion in each ieraion. Signal oupu is again processed by applying ino algorihm, where produc o modiied gain uncion and noisy speech specrum is used o give enhanced speech signal. Sysem Model Assuming speech and noise are uncorrelaed, noisy speech signal x(n) can be represened as ( n) s( n) d( n) x + (4) where s(n) is clean speech signal and d(n) is noise signal. Signal is divided ino overlapped rames by window and shor-ime Fourier ransorm (STFT) is applied o each rame. Power specrum o noisy speech X ( can be represened as ( k ) S ( k ) + D ( l k ) (5) X, where (k,... ) are requency bin index and (l,,..., ) are rame index. S( is power specrum o clean speech signal and D( is power specrum o noise signal. Proposed algorihm is summarized in low diagram (Fig. ). Proposed Mehod Noise reducion algorihm is based on variance o noisy power specrum (NPS) in a ime and requency dependen manner as x X ( k ) and k ( k ) l X ( l, k ) (6) where x (l) is average NPS in requency bin, and x ( is average NPS or rame index. x

3 334 J SCI IND RES VO 70 MAY 0 ( ) ν ( k ) ( ) ( ) X l, k x k l and ( ) ν ( ) k X ( l, k ) x ( l ) (7) l Which represens variance o noisy specrum in ime and requency bins. σ v and σ k v ( (8) l v v γ and γ σ σ (9) where σ and σ are average esimae o noise power. Eq. (9) gives raio o variance or NPS in ime-requency (T-F) bin o is average. In a region, where a speech signal is srong, variance raio will be high. Thereore, Eq. (9) is used o deermine speech-presence or speechabsence in T-F bins. Classiicaion o Speech-Presence and Speech-Absence in Frames Using an Adapive Sigmoid Funcion Proposed mehod uses an adapive algorihm 6-8 wih a sigmoid uncion o rack hreshold and o conrol speech disorion and residual noise as µ ( l) + exp(0 *( γ ( l) ( l))) (0) where µ (l) is adapive hreshold, is conrol parameer. Threshold µ (l) is varied based on, which is derived rom linear uncion using a poseriori in rame index. s ( ) o + () ( l ) norm ( X ( l, k ),) 0 log norm ( ) D l, () 5 5 where ( ) ( ) D l X l, k is mean value o l X ( l, k ) or iniial 5 rames during he period o irs silence and norm is Euclidean lengh o a vecor. o. (3) max s min min max s (4) max min where s is slope o, o is ose. Consans ( min 0., max 0.5, min -5 db and max 0 db) are experimenal oucomes obained by rail and error mehod o ge an opimum value. Poseriori in Eq. () conrols. An increase in is good or noisy signals wih a low (< 5 db), and a decrease in is good or relaively high (> 5 db). Thus, conrol radeo beween speech disorion and residual noise in rame index using. I a speech signal is presen, µ (l) is calculaed by Eq. (0), which is exremely small (very close o 0), oherwise value o µ (l) will be approximaely. Updaed Noise Power Specrum Using Classiicaion o Speech-Presence and Absence in Frames Classiicaion rule o deermine speech presence / absence in a rame is based on algorihm 8, 9 as i ( µ ( ) > δ ) (Speech absen rame) l D noise( l +, λ * D noise( + ( λ) X( ( k ) G( k )τ. G up else (Speech presen rame) l l k D mean, ( l ) D noise ( l k ) D noise ( D k m k mean and ( k ) G( l, k )(. τ ) G up (5) where decision parameer δ, scaling acor τ are iniially 0.99 and gain uncion G( is.0. Threshold µ (l) is compared wih δ. I i is greaer han δ, hen speech is

4 DEEPA & SHANMUGAM: ENHANCEMENT OF NOISY SPEECH SIGNA FOR DIGITA HEARING AID 335 deermined o be absen in l h rame; hen, noise is D mean updaed using recursive equaion and ( l ) is calculaed by averaging noise over rames. Value λ is a smoohing consan where value is o be se appropriaely beween 0 and. Oen a consan value o 0.85 o 0.95 is suggesed. D mean ( l ) is assumed esimae o residual noise o speech rame. G up ( is updaed gain uncion in a rame index using gain uncion G( and τ or rames, in which speech is absen. I l h rame is D mean considered o be speech presen rame, ( l ) D noise. D mean o ( o rames in presence o speech as X, upd is se is used o reduce sicky noise ( l, k ) X ( k ) D noise ( l k ) (6) Updaed NPS o rame index is dierence beween NPS and esimaed noise. Calculaion o Adapive Threshold Based on Sigmoid Funcion in Frequency Bins In a manner parallel o bins, an adapive algorihm wih a sigmoid uncion 0 is used o rack hreshold in a requency bin as µ ( + exp( 0*( γ ( ( )) (7) where µ ( is adapive hreshold using sigmoid uncion in requency bins. Threshold µ ( is adapive in he sense ha i changes depending on conrol parameer. Conrol parameer o requency bin is derived rom linear uncion using a poseriori in requency bins as ( k ) s ( k ) o. + (8) ( 0.log D X ( k ) noise ( ) k (9) where ( ) D noise k is esimae o NPS in requency bins. o s s min max min max min (0) where s is slope o, o is ose o. Consans ( min 0., max 05, min - 5dB and max 0dB) are experimenal values obained by rail and error mehod o ge an opimum value. Noise Reducion Using a Modiied Gain Funcion and Updaed Noisy Power Classiicaion algorihm or deermining speech presence or absence in a requency bin is G I ( ( k ) δ ( k ) G ( k )τ. mi µ > ) (Speech absen rame) upd else (Speech presen rame) ( k ) G ( k )(. τ ) Gmi upd () where decision parameer δ is iniially I µ ( is greaer han δ, hen speech is deermined o be absen in l h requency bin. G mi ( represens modiied gain uncion or T-F bins using gain uncion G upd ( and scaling acor τ as S( Gmi( * X upd( () Esimaed enhanced speech power specrum can be obained rom he produc o modiied gain uncion or T-F bins and updaed NPS o T-F bins. Enhanced speech specrum can hen be ransormed back o ime domain using inverse STFT and synhesis wih overlap-add mehod. Objecive Measures or Perormance Evaluaion Objecive measures are based on a mahemaical comparison o original and processed speech signals. I is desired ha objecive measures be consisen wih judgmen o human percepion o speech. and mean

5 336 J SCI IND RES VO 70 MAY 0 square error (MSE) are wo o he mos widely used objecive measures. Signal-o-Noise Raio () is calculaed as raio o signal o noise power in decibels. I summaion is perormed over whole signal lengh, operaion is called global PSD comparison Noisy Signal Enhanced Signal Clean Signal PSD db S 0log 0 n (n) n [ S(n) Ŝ(n) ] (3) where S(n) is noisy speech signal and S(n) is enhanced signal. Mean Square Error (MSE) MSE measures average o square o error as MSE ( Soriginal( n) S( n)) (4) n 0 where S original (n) is clean signal. Subjecive Measures or Perormance Evaluaion Subjecive qualiy raings were obained using ITU- TP.835 mehodology designed o evaluae speech qualiy along hree dimensions (signal disorion, noise disorion and overall qualiy). Improvemens o PESQ measure are repored along wih new composie objecive measures, as: A) speech signal alone using a ive-poin scale o signal disorion (SIG) [5-very naura no degradaion, 4-airly naura lile degradaion, 3- somewha naura somewha degraded, -airly unnaura airly degraded, -very unnaura very degraded]; B) background noise alone using a ive-poin scale o background inrusiveness(ba) [5-no noiceable, 4-somewha noiceable, 3-noiceable bu no inrusive, -airly conspicuous, somewha inrusive, - very conspicuous, very inrusive]; and C) overall eec using scale o mean opinion score (OVR) [-bad, - poor, 3-air, 4-good, 5-excellen]. Resuls and Discussion Tes samples were aken rom SpEAR (speech enhancemen assessmen resource) daabase o CSU Frequency-----> Fig. PSD plo o car noise, clean and enhanced signal Table Comparison o value o proposed mehod wih exising mehods Type o noisy signal DEF SS Proposed mehod mehod mehod Whie Saionary (WS), 0 db Whie bursing (WB), 0 db Whie bursing, 3 db Car phone Cellular Facory phone Pink saionary, 0 db Pink saionary, 6 db Table Comparison o subjecive measures Noisy Convenional mehod 0 Proposed mehod signal CSIG CBA COV CSIG CBA COV WS, 0 db WB, 0 db WB, 3 db Car Cellular (Cener or Spoken anguage Undersanding). Power specral densiy (PSD) o reerence clean signa noisy signal and enhanced signals or car noisy signal were obained and compared or car noisy signal (Fig. ). Oupu shows ha noise signal in higher requency range (40-80) is minimized. calculaed by proposed mehod gave very high values (Table ) as compared o DEF (Dual Exended alman Filering) and SS (Specral Subracion) mehods. Also, subjecive qualiy raings by proposed mehod were beer han convenional mehod (Table ). MSE by proposed mehod was ound minimum

6 DEEPA & SHANMUGAM: ENHANCEMENT OF NOISY SPEECH SIGNA FOR DIGITA HEARING AID 337 MSE Whie Saionary 0dB Whie bursing 0dB Whie bursing 3dB Car phone Type o signal Fig. 3 Mean square error (MSE) comparison (Fig. 3) when compared wih convenional mehod. Thus proposed mehod perorms ar beer han convenional mehod. Conclusions Proposed mehod used variance and modiied gain uncion wih PDE preprocessing. I improved speech qualiy by improving (> 0 db), and can be accommodaed in digial hearing aids where 5-5 db raise is required or sensorineural loss paiens. MSE value is minimum in proposed mehod. This approach is more suiable or boh saionary and non saionary noise environmens, and can be implemened in digial signal processor (TMS 30C673) or processing real ime signals. Proposed mehod ouperorms sandard power specral subracion mehod and mehod using sandard deviaion, resuling in superior speech qualiy and largely reduced musical noise in single channel sysem or boh saionary and non saionary noise environmens. Cellular MSE o o Convenional mehod MSE o o Proposed mehod Reerences Virag N, Single channel speech enhancemcn based on masking properies o he human audiory sysem, IEEE Trans Speech Audio Process, 7 (999) 6-37 Boll S F, Suppression o acousic noise in speech signal specral subracion, IEEE Trans Acous Speech Signal Process, 7 (979) Ephraim Y & Malah D, Speech enhancemen using a minimum mean-square error log-specral ampliude esimaor, IEEE Trans Acous Speech Signal Process, ASSP-3 (984) Aya S S, Manzuri M, Diana R & abudian J, An improved specral subracion speech enhancemen sysem by using an adapive specral esimaor, in 005 IEEE Canadian Con on Elecrical & Compu Engg (Canada) May 005, Won Seok J & Sung Bae, Reducion o musical noise in specral subracion mehod using subrame phase randomisaion, IEEE Elecron e, 35 (999) l Sundarrajan R & Philipos C, A noise esimaion algorihm or highly non- saionary environmens, Speech Commun, 48 (006) Cohen I, Noise specrum esimaion in adverse environmens: improved minima conrolled recursive averaging, IEEE Trans Speech Audio Process, (003) Rangachari S, oizou P & Hu Y, A noise esimaion algorihm wih rapid adapaion or highly nonsaionary environmens, in Proc ICASSP-004 (Monrea Canada) May 004, ee S J & im S H, Speech enhancemen using gain uncion o noisy power esimaes and linear regression, in Proc IEEE/FBIT In Con Froniers in Convergence o Bioscience & Inorm Techno Ocober 007, ee S-J & im S-H, Noise esimaion based on sandard deviaion and sigmoid uncion using a poseriori signal o noise raio in nonsaionary noisy environmens, In J Con, Auoma & Sys, 6. (008) Quackenbush S, Barnwell T & Clemens M, Objecive Measures or Speech Qualiy Tesing (Prenice-Hal Englewood Clis) 988. Ma J, Hu Y & oizoub P C, Objecive measures or predicing speech inelligibiliy in noisy condiions based on new band-imporance uncions, J Acous Soc Am, 5 (009)