Robust speech recognition using harmonic features

Transcription

1 Published in IE Signal Processing Received on 29h Sepember 2012 Revised on 14h July 2013 Acceped on 22nd July 2013 ISSN Robus speech recogniion using harmonic feaures Yeh Huann Goh, Paramesran Raveendran, Sudhanshu Shekhar Jamuar Faculy of Engineering, Deparmen of Elecrical Engineering, Universiy of Malaya, Lembah Panai, Kuala Lumpur 50603, Malaysia Absrac: In his sudy, he auhors propose a speech recogniion sysem using harmonic srucure relaed informaion o deec harmonic feaures in noisy environmen. he proposed algorihm firs exracs he harmonic componens conained inside he speech signals using sine funcion convoluion. By seing he frequency of he sine funcion as equal o he fundamenal frequency of speech signals, harmonic componens can be exraced ou. he reconsruced signal obained by summing up he exraced harmonic componens is found o have a high degree of correlaion wih he original signal. he exraced frame energy measure of he harmonic componens has been furher processed o become dynamic harmonic feaures and hen used ogeher wih he European elecommunicaions Sandards Insiue ESI) fron-end processed mel-frequency cepsral coefficiens MFCC) feaure or he percepual linear predicion PLP) feaure in he speech recogniion sysem. he proposed enhanced speech recogniion sysem shows a beer recogniion rae over he ESI fron-end processed MFCC or PLP)-based speech recogniion sysem. 1 Inroducion Speech is one of he primary sources of human communicaion. However human communicaion is no always face-o-face or human-o-human. here are several siuaions such as auomaic ranscripion, mulimedia conen analysis, and naural human compuer inerfaces in which mechanical recogniion of speech is necessary. he primary hrus in speech recogniion herefore has been he search for a meric, robus o environmenal noise, by which he speech can be recognised accuraely in real ime, and allows speech-based applicaions such as speech-o-ex ranslaion. Voiced speech signals are buil from harmonic componens which are periodic and can be resynhesized easily. Many algorihms o exrac harmonic componens from speech signals have been proposed where he moivaion for he inensive research is is wide usabiliy wihin he field of engineering such as: speech analysis [1 4], speech coding [5 7], pich racking and esimaion [3, 7, 8], speech enhancemen using harmonic regeneraion [9], pos-speech enhancemen [10, 11], bandwidh exension [12, 13] and voice aciviy deecion [14]. Recenly, he periodiciy of harmonic srucure has been used o compensae noise and o idenify he corruped ime-frequency regions. A new mehod for he enhancemen of speech in which argeed speaker enrollmen as well as sysem raining wihin he ypical noise environmen are feasible [15]. A new clean signal based on is likely characerisics esimaed from he disored signals has been resynhesized insead of filering. he resuling signals show an increased nauralness and percepual qualiy in speech. Anoher new applicaion of he harmonic model has been an add-on module wih several popular noise reducion mehods [4]. he add-on module uilises he harmonic plus noise model HNM) of speech o rerieve damaged speech srucure. An improved sinusoidal modelling mehod based on percepual maching pursuis compued in he bark scale has been proposed by [16] for parameric audio coding applicaions. Besides, a long-shor frame associaed harmonic LSAH) model [17] has been used o separae wo speech sources from a single-channel recording. he human being has he abiliy o perceive imporan harmonic componens even in very noisy environmens [18]. Such, however, is no he case wih sysems devised o recognise speech. Algorihms based on harmonic srucure relaed HSR) informaion have been developed under low signal-o-noise raio SNR) condiions o mimic he human capabiliy o deec harmonic feaures in noisy environmens. Mos researchers use comb filering echnique o exrac he harmonic componens from he speech signals [19 24]. here are wo ypes of comb filering: finie impulse response FIR) filering and infinie impulse response IIR) filering. he FIR comb filer causes emporal blurring effecs on quasiperiodic speech. IIR comb filering ha cascades a number of second-order IIR band pass filers o creae a high-order comb filer improves upon he undesirable characerisics of he FIR comb filer. However, he high-order IIR comb filers are ofen no feasible because of insabiliy consrains [25]. In his paper, we propose a sine funcion convoluion mehod o exrac he harmonic componens ou from voiced speech signals. Proposed convoluion mehod do no suffer from insabiliy problem. Mahemaically, only he inended harmonic componens are exraced ou and all oher harmonic componens have a gain of zero afer he IE Signal Process., pp & he Insiuion of Engineering and echnology 2013

2 convoluion process. he proposed convoluion mehod has comb-filer like frequency response wih one excepion: here is no aenuaion a he desired oupu frequency. o exrac ou he desired harmonic componen, he period of he sine funcion has been se o equal o he pich period of ha paricular harmonic componen. As he order of he harmonic componen ges higher, pich period ges shorer. Since he sampling rae of he speech daa is 8 khz, he number of harmonic componens o exrac ou is needed o be conrol wihin a cerain range so ha a proper sine funcion can be formed using limied number of daa poins. Wih more number of daa poins o form he sine funcion, he more complee harmonic componen can be exraced ou. As a resul, he number of harmonic componens necessary o represen he voiced speech signals has been found using previous sudy by Schwarz e al. [26] and reconsrucion es. In order o validae his algorihm, IDIGI and IMI speech daa ses have been used for raining and esing. he IDIGI daa se consiss of more han isolaed and sequence digis for every digi beween 1 and 9 and wo differen uerances of he digi 0, i.e. zero and oh) spoken by 326 men, and women from coninenal US. he age range of he speakers is from 17 o 70 years old. he IMI daa se conains broadband recordings of 630 speakers of eigh major dialecs of American English, each reading en phoneically rich senences. I is a corpus of phonemically and lexically ranscribed speech of speakers of differen sexes and dialecs. esing and raining subses are available in boh IDIGI and IMI daabases for esing and raining purposes, respecively. An algorihm o exrac he speech signals harmonic componens has been developed. he exraced harmonics have been used o reconsruc he original voiced signals and he reconsruced voiced signal has been compared wih he original voiced signal using he Pearson correlaion coefficien. Comparisons have been made beween European elecommunicaions Sandards Insiue ESI), he fron-end feaure processing based on mel-frequency cepsral coefficien MFCC), percepual linear predicion PLP) or RASA-PLP feaures [27] and he exraced dynamic harmonic log-energy level concaenae wih he same MFCC, PLP or RASA-PLP feaures. he resul shows ha he proposed algorihm has a beer recogniion rae in noisy environmen. he organisaion of he paper is as follows: Secion 2 begins wih a brief review of he harmonic model of voiced speech signals followed by he exracion of harmonic componens, by convolving he sine funcion wih voiced speech signals. In Secion 3, we show how exraced harmonic componens are reconsruced back and see how hey correlae wih he original voiced speech signals. Secion 4 furher processes he exraced harmonic componens ino dynamic harmonic feaures DHF) and concaenae i wih he MFCC, PLP or RASA-PLP feaures in speech recogniion sysem. Secion 5 presens he conclusions and recommendaions for fuure work of his sudy. 2 Exracion of energy-based harmonic feaures 2.1 Harmonic model of voiced speech signal A periodic voiced speech signal consiss of a fundamenal one and several consiuen ones whose frequencies are higher han ha of he fundamenal noe. he frequency of harmonic ones is N imes he frequency of he fundamenal one, where N is a posiive ineger. he harmonic model of a voiced frame, V) can be represened in he following way V) = b c0 + N b cn cos 2np ) + b sn sin 2np )) + N ) 1) where b c0 is he mean of he voiced speech signal, coefficiens b cn and b sn carry he informaion of he inensiy and phase of he nh harmonic of he voiced speech signal, N denoes he number of harmonics, is he pich period of he signal and N ) is he Gaussian noise variable a ime [28]. We can rewrie 1) as follows [ [ V ) = N A n, sin 2np ]] + k n, + N ) 2) where coefficien A n, represens he ampliude and k n, is he phase shif of he nh-order harmonic one a ime and he mean of he voiced speech signal b c0 is zero. hus b cn = A n, sin2nπ/ )k n, ), b sn = A n, cos2nπ/ )k n, ). Alhough speech signals are non-saionary signals and heir saisical properies vary wih ime, heir variaion is slow in a shor period of ime. Speech signals can be assumed as saionary signals in he shor inerval. hus 2) can be represened as follows [ [ V ) = N A n sin 2np ]] + k n + N ) 3) 2.2 Sine funcion convoluion Convoluion is a mahemaical operaion on wo funcions f and V, producing a hird funcion ha is ypically viewed as a modified version of one of he original funcions. In he proposed mehod, funcion V represens periodic voiced speech signals 1), and funcion f represens he sine wave funcion, B represens he period of he sine wave funcion sin 2mp ) f s = B s 0 s, B, m = 1, 2, 3,... 0 oherwise We se = 0 s, where 0 is he saring ime of he convoluion process, he convoluion of he voiced speech signal funcion 3) Vf)) is given as 1 VH 0, B = = 1 B f s )V 0 s ds 4) sin 2mp ) 1 s A 0 B n [ [ sin 2np ] ] 0 s + k n + N 0 s d s For discree-ime speech signals, he convoluion is given by 5) 2 & he Insiuion of Engineering and echnology 2013 IE Signal Process., pp. 1 9

3 B [ ] VH 0, i = f i)v 0 i i=0 he sine wave convoluion has a comb-filer like frequency response wih one excepion ha here is no aenuaion when he frequency componen of he process signal is equivalen o he frequency of he sine wave. his excepion makes he sine funcion convoluion suiable o be used o exrac ou all he individual harmonic componens. We obain he expression for VH 0, B) from 5) by inegraing i over ime 0 o B. he frequency conen of noise N 0 s ) in 5) is much higher han he range of fundamenal frequencies of any male or female speaker. herefore he noise signal ges filered ou by he sine funcion convoluion band pass filer, and he expeced value EN 0 s )) is relaively low in ampliude. he deailed seps are given in Appendix 1. N VH 0, B = 6) ma n 2 B cos np p m 2 2 n 2 B 2 B + 2 )) 0 + 2k n sin npb 2.3 Order of harmonic wih maximal power in human speech Schwarz e al. [26] have shown ha he sound wave has maximum power a he fundamenal frequency of he laryngeal oscillaion, and a rich se of overones a frequency values approximaing ineger muliples of he fundamenal and he power in he specrum of he waveform decreases exponenially wih increasing frequency. For simpliciy, Schwarz e al. have considered he vocal rac as an ideal pipe closed a one end measuring 17 cm, which is approximaely he lengh of he vocal rac in adul human males. he firs resonance frequency, called he forman occurs a 500 Hz and he power of he laryngeal pressure is leas aenuaed a his frequency. hus he specral componen of he waveform having maximum power in adul human uerances will be close o 500 Hz. he voiced speech of a ypical adul male will have a fundamenal frequency of from 85 o 155 Hz, and ha of a ypical adul female from 165 o 255 Hz [29 31]. Using he 500 Hz resonance frequency saed above, he frequency of he harmonic ha should mos ofen be he frequency a which he power is maximal of an adul male speaker is from he hird o he sixh harmonic, and for a female speaker, he second and hird harmonic. he empirical disribuion of ampliude maxima ploed according o harmonic number for he speech sound specra in he IMI corpus is shown in Fig. 1. I can be seen from he figure ha more han 90% of he ampliude maxima occur a harmonics numbered 1 8 [26]. As a resul, he harmonic number, N necessary o be used in speech processing based on he proposed DHF has been se o eigh in his paper. 7) voiced. In his paper, he wo firs classes are grouped ino one single unvoiced class, and he oher wo ino one single voiced class. he ESI pich exracor gives an oupu pich value of 0 if a paricular speech frame has been classified as unvoiced class. 2.5 Using sine funcion convoluion o exrac he harmonic componens We now use he sine funcion convoluion o exrac he harmonic componens. We se he value of B = in sine funcion f s ) equal o he deeced pich period from Secion 2.2 in 7). he ineger m can be any value and ime 0 is an independen variable which varies from begin ime o he end ime of he speech signal. When m n, we obain VH 0, = 0 8) Because of he sinnπb/ ) componen in 7) and for m = n,we obain A VH 0, = m 2 cos 2pm 0 + k m using he squeeze heorem [32]. he derivaion of 7) 9) is shown in Appendix 2. From 3), since he noise signal has been filered ou, he harmonic componens of voiced speech signals is represened as [ HC m, 0 = Am sin 2mp ] 0 + k m 9) 10) he value m in 10) represens he order of harmonic componens of voiced speech signals in 11) 2 HC m, 0 = VH 0 + ) 4m, 11) Fig. 2 shows he plo of harmonic componens HC from 1s order o 8h order for unsegmened speech signals. Fig. 3 shows he segmened speech signal one spoken by a female speaker. Figs. 3b i clearly show ha all he processed signals look like sine funcions wih funcion period, /2,.../8, where denoes he pich period of he speech signal. 2.4 Pich period deecion he ESI pich esimaion algorihm [27] is used o deermine he pich informaion in each frame. he ESI pich deecion algorihm classifies all he speech frames ino four classes: i) non-speech, ii) unvoiced, iii) mixed-voiced and iv) fully Fig. 1 Probabiliy disribuion of he harmonic number a which he maximum ampliude occurs in speech sound specra derived from he IMI corpus IE Signal Process., pp & he Insiuion of Engineering and echnology 2013

4 Fig. 2 Unsegmened speech signal one spoken by female speaker and is corresponding 1s o 8h harmonic componens a Original speech from 300 o 900 ms b Firs harmonic from 300 o 900 ms c Second harmonic from 300 o 900 ms d hird harmonic from 300 o 900 ms e Fourh harmonic from 300 o 900 ms f Fifh harmonic from 300 o 900 ms g Sixh harmonic from 300 o 900 ms h Sevenh harmonic from 300 o 900 ms i Eighh harmonic from 300 o 900 ms 4 & he Insiuion of Engineering and echnology 2013 IE Signal Process., pp. 1 9

5 Fig. 3 Segmened speech signal one spoken by female speaker and is corresponding 1s o 8h harmonic componens a Original speech from 600 o 620 ms b Firs harmonic from 600 o 620 ms c Second harmonic from 600 o 620 ms d hird harmonic from 600 o 620 ms e Fourh harmonic from 600 o 620 ms f Fifh harmonic from 600 o 620 ms g Sixh harmonic from 600 o 620 ms h Sevenh harmonic from 600 o 620 ms i Eighh harmonic from 600 o 620 ms IE Signal Process., pp & he Insiuion of Engineering and echnology 2013

6 3 Reconsrucion es he harmonic componens obained for a given speech signal compued from 11) has been used in reconsrucion of speech using he following equaion V) = number of componens m=1 HCm, ) 12) he reconsruced voiced signal V) has been compared wih he original voiced speech signal V ) using Pearson correlaion coefficien given by r V, V = covv, V) d V d V [] = E V m V V m V 13) where μ V and m V are expeced value of V and V, δ V and d V are sandard deviaions of V and V. he reconsrucion es has been carried ou over all he clean senences conained inside he IDIGI daabase raining subse. he es was carried wih harmonic componens ranging from 1 o 8. he correlaion rv, V )) beween original and reconsruced voiced speech signals using differen number of harmonic componens is shown in Fig. 4. he correlaion is bes by using eigh componens and has a value he correlaion drops as he number of harmonic componens is reduced. he degradaion is slow ill four harmonic componen are used. Below four, he correlaion drops sharply. I can be concluded ha HCm, ) represens he harmonic signals of voiced speech signal and ha he error beween reconsruced speech signal and he original speech signal reduces significanly for higher harmonic componens. Fig. 5 shows he plo of original and voiced speech signals reconsruced by using eigh harmonic componens compued in 11). Reconsruced voiced signals using five or more harmonics are highly correlaed o he original voiced speech signals wih he achieved correlaions exceeding 0.9. hese resuls are in agreemen wih he resuls obained by Schwarz e al. [26] ha more han 75% of he ampliude maxima occur a harmonic numbers 2 5 as shown in Fig 1). Fig. 5 Demonsraion of original and is corresponding reconsruced voiced speech signals using eigh harmonic componens for speech signal one, spoken by a female speaker 4 Experimenal sudy A number of experimens were carried ou wih he IMI and IDIGI daabases using he following procedure: each speech signal is firs down sampled o 8 khz. hen, direc curren DC) offse of he inpu speech signal is removed. he offse-free inpu signal is hen divided ino overlapping frames. he frame lengh is 25 ms and he frame shif inerval is 10 ms. A pre-emphasis filer is applied o he framed offse-free inpu signal. his follows by applying a Hamming window of lengh 25 ms o he oupu of pre-emphasis block. Pich informaion of each frame is esimaed using he ESI pich esimaion algorihm. Equaion 6) is used o exrac he harmonic componens from he Hamming window processed frame. All he speech frames are processed using 6) wih differen pich period,. If a paricular speech frame is classified as mixed-voiced or fully voiced, deeced pich informaion will be used as pich period,. If i is being classified as non-speech or unvoiced, pich period, is fixed as he average of all he deeced pich period conain inside ha paricular speech signal. Harmonic componens in each frame are exraced. Logarihmic frame energy measures are compued afer he exracion of harmonic componens in each frame. he logarihmic frame energy measure of each exraced harmonic componen frame is compued using he following equaion EHCx, m) = ln M ) HCm, ) 2 =1 14) where EHC represens he logarihmic frame energy measure of he xh frame, and he mh-order harmonic componen of M samples. A floor is used in he energy calculaion, which makes sure ha he resul for EHC is no less han 1.6. his follows by he exracion of he dynamic harmonic feaures, DHF using he following equaion DHFm) = EHCx + 1, m) EHCx 1, m) 2 15) Fig. 4 Correlaion beween original and reconsruced voiced speech signals using differen numbers of harmonic componens he exraced DHF are furher concaenaed wih he MFCC, PLP or RASA-PLP feaures o creae combinaion feaures. In he proposed feaure combinaion speech recogniion sysem, 12 MFCC, PLP or RASA-PLP feaures, 1 energy level, 12 dela MFCC, PLP or RASA-PLP feaures, 1 dela energy level and differen number of dela logarihmic energy of harmonic componens are used in raining and esing. Fig. 6 is a block diagram of he feaures exracion 6 & he Insiuion of Engineering and echnology 2013 IE Signal Process., pp. 1 9

7 Fig. 6 Block diagram of he MFCC-feaures and DHF exracion algorihm algorihm. HK-oolki was used in building hidden Markov models HMM). 4.1 IDIGI daabase For experimens ha were carried ou wih he IDIGI daabase, HMM-based speech recogniion sysems have been developed. All hese HMM models consis of eleven 16-sae coninuous words excep silence and pause, ha have 3 and 1 saes, respecively), wih four Gaussians per sae. raining was done using 8598 clean senences conained inside he raining subse of he daabase and ess were carried ou over 8700 senences conained inside he esing subse of he daabase wih addiive whie noise a differen SNRs clean, 20, 15, 10, 5, 0 and 5 ). able 1 shows he recogniion raes of differen ypes of speech recogniion sysem. For MFCC plus DHF-based speech recogniion sysems, all hree ypes of combinaion feaures-based speech recogniion sysems give a beer recogniion rae a all SNRs. he recogniion rae a high SNR regions clean, 20, 15 and 10 ) are nearly he same for all hree ypes of speech recogniion sysems using a differen number of dela harmonic feaures. As for low SNR regions 5, 0 and 5 ), he speech recogniion sysem using eigh dela harmonic feaures shows he bes recogniion rae. For PLP-based speech recogniion sysems, recogniion raes for speech recogniion sysem wih or wihou he DHF feaures are nearly he same a high SNR regions clean, 20 and 15 ), significan improvemen can be observed a low SNR regions 10, 5 and 0 ). As for he RASA-PLP R-PLP)-based speech recogniion sysem, only sligh improvemen can be seen in he combinaion feaures-based speech recogniion sysem for all SNRs. his able 1 Word accuracy WAcc) of differen ypes of feaurebased speech recogniion sysems Clean MFCC ΔMFCC 13MFCC ΔMFCC + 6DHF 13MFCC ΔMFCC + 8DHF 13MFCC ΔMFCC + 10DHF 13PLP + 13ΔPLP PLP ΔPLP + 8DHF 13R PLP ΔR PLP 13R PLP + 13ΔR PLP + 8DHF is mainly caused by he effeciveness of he RASA filering echnique o filer he speech signal componens which have slow frame-o-frame specral changes. In his experimenal sudy, addiive whie Gaussian noise which is a random signal wih a fla power specral densiy was added ino he speech signals. Afer he RASA filering process, mos of he whie Gaussian noise has been aenuaed and leaving us wih he speech signals componens which have faser channel variaion including he harmonic componens. Overall, he MFCC plus DHF speech recogniion sysem shows he bes recogniion rae over clean speech signals, and RASA-PLP plus DHF speech recogniion sysem shows bes robusness. 4.2 IMI daabase For experimens ha were carried ou wih he IMI daabase, phoneme HMM-based speech recogniion sysems using bigram language model were developed. Developed phoneme HMMs are hree-sae, lef righ wih eigh Gaussian mixures. raining was done wih he clean senences conained inside he raining subse and ess were carried ou over senences conained inside he es subse wih addiive whie noise a differen SNRs clean, 20, and 5 ). able 2 shows he correc phoneme accuracies of differen ypes of speech recogniion sysems. es recogniion raes obained over he IMI daabase show he same rend as he es recogniion raes obained wih he IDIGI daabase. By concaenaing he DHF ino he MFCC-based speech recogniion sysem, he recogniion rae improves significanly a all SNRs. However, only sligh improvemen can be seen by concaenaing he DHF ino he RASA-PLP-based speech recogniion sysem a all SNRs. Again, he MFCC plus DHF-based speech recogniion sysem shows he bes recogniion rae over clean speech signals, and he RASA-PLP plus DHF-based speech recogniion sysem shows bes robusness. 4.3 Effecs of he accuracy in pich esimaion on he speech recogniion rae o exrac ou he proposed DHF, pich inerval of each speech frame is compued and used as he period of he able 2 Correc phoneme accuracies %) of differen ypes of speech recogniion sysems Clean MFCC + 13ΔMFCC MFCC + 13ΔMFCC DHF 13R PLP + 13ΔR PLP R PLP + 13ΔR PLP + 8DHF IE Signal Process., pp & he Insiuion of Engineering and echnology 2013

8 able 3 Effecs of he false pich inerval deecion on he WAcc of he speech recogniion sysem Sandard deviaion, ms WAcc sine funcion convoluion. A false deeced pich inerval will cause error in he compuaion of he proposed DHF. o es he effecs of he false deeced pich inerval on he performance of he proposed mehod, he same speech recogniion sysem as before using 12 MFCC-feaures, 1 energy level, 12 dela MFCC-feaures, one dela energy level and eigh DHF wih addiive whie noise a SNR level 15 was used. he es was carried ou wih he IDIGI speech daabase bu he deeced pich inerval, using ESI pich esimaion algorihm in each speech frame was alered by adding normally disribued variable R wih mean 0 ms and differen sandard deviaions 0.00, 0.25, 0.50, 0.75 and 1.00 ms) using 16). he speech recogniion sysem using he normally disribued variable R wih sandard deviaion 0 ms acs as he reference model since he deeced pich inerval was no o be alered. Wih oher speech recogniion sysems using variable R wih larger sandard deviaions, however, larger error in he pich period deecion happened hus causing greaer effecs on he proposed feaures = + R 16) able 3 shows he differen recogniion raes of differen speech recogniion sysems using false alered pich inervals. Resuls show ha for pich inervals alered by normal disribued variable R wih low sandard deviaions 0.25 and 0.50 ms), recogniion rae drops a bi. Referring back o able 1, i can be seen ha recogniion rae for he MFCC-based speech recogniion sysem is 86.5%. his shows ha he proposed speech recogniion sysem using DHF wih small error in he deeced pich inerval can sill provide improvemen in he recogniion rae. However, once he error in he pich inerval obains larger by using he normally disribued variable R wih large sandard deviaions 0.75 and 1.00 ms), recogniion rae drops significanly. hese recogniion raes are even lower han he recogniion rae achieved by he convenional MFCC-based speech recogniion sysem under same condiions. 5 Conclusion An algorihm o exrac he harmonic componens using sine funcion convoluion has been developed. he speech signal is reconsruced by summing up he exraced harmonic componens and he reconsruced signal is found o have a high degree of correlaion wih he original signal. I has been shown ha he recogniion rae of he proposed enhanced speech recogniion sysem using combinaion feaures is beer han he ESI fron-end processed MFCC, PLP or RASA-PLP-based speech recogniion sysem. Alhough harmonic exracion using he harmonic model has proven successful in improving he speech recogniion rae of he proposed algorihm, since he harmonic model can jus represens he voiced speech signals, oher speech models, such as he HNM model, which can represen boh voiced and unvoiced speech signals should be used in fuure work. 6 Acknowledgmen his work was suppored by he HIR-MOHE gran no. UM.C/HIR/MOHE/ENG/42. 7 References 1 Irizarry, R.: he addiive sinusoidal plus residual model: A saisical analysis. Proc. CNMA, Park, S., Kwon, W., Kwon, O., Kim, M.: Shor-ime Fourier analysis via opimal harmonic FIR filers, IEEE rans. Signal Process., 2002, 45, pp abrikian, J., Dubnov, S., Dickalov, Y.: Maximum a-poseriori probabiliy pich racking in noisy environmens using harmonic model, IEEE rans. Speech Audio Process., 2004, 12, pp Zavarehei, E., Vaseghi, S.: Inerpolaion of los speech segmens using lp-hnm model wih codebook pos-processing, IEEE rans. Mulimedia, 2008, 10, pp Griffin, D., Lim, J.: Muliband exciaion vocoder, IEEE rans. Acous., Speech Signal Process., 2002, 36, pp Kondoz, A.: Digial speech: coding for low bi rae communicaion sysems John Wiley & Sons Inc, 2004) 7 Codec, D.: Version 2, Inmarsa-M Specificaion, Inmarsa, Gong, Y., Haon, J.: ime domain harmonic maching pich esimaion using ime dependen speech modeling, IEEE rans. Acous., Speech Signal Process., 2003, 35, pp Plapous, C., Marro, C., Scalar, P.: Speech enhancemen using harmonic regeneraion. IEEE In. Conf. on Acousics, Speech, and Signal Processing, 2005, vol. 1, pp Yu, A., Wang, H.: New speech harmonic srucure measure and i applicaion o pos speech enhancemen, IEEE In. Conf. Acous., Speech, Signal Process., 2004, 1, pp. I Zavarehei, E., Vaseghi, S., Yan, Q.: Noisy speech enhancemen using harmonic noise model and codebook-based pos-processing, IEEE rans. Audio, Speech, Lang. Process., 2007, 15, pp Vaseghi, S., Zavarehei, E., Yan, Q.: Speech bandwidh exension: exrapolaions of specral envelop and harmoniciy qualiy of exciaion, IEEE In. Conf. Acous., Speech Signal Process., 2006, 3, pp. III Raza, D., Chan, C.: Enhancing qualiy of celp coded speech via wideband exension by using voicing GMM inerpolaion and HNM re-synhesis. IEEE In. Conf. on Acousics, Speech, and Signal Processing, IEEE, 2005, vol Fukuda,., Ichikawa, O., Nishimura, M.: Long-erm specro-emporal and saic harmonic feaures for voice aciviy deecion, IEEE J. Sel. op. Signal Process., 2010, 4, pp Xiao, X., Nickel, R.: Speech enhancemen wih invenory syle speech resynhesis, IEEE rans. Audio, Speech, Lang. Process., 2010, 18, pp Vera-Candeas, P., Ruiz-Reyes, N., López-Ferreras, F.: Bark scale-based percepual maching pursui for improving sinusoidal audio modeling, Digi. Signal Process., 2009, 19, pp Huang, Q., Wang, D.: Single-channel speech separaion based on long-shor frame associaed harmonic model, Digi. Signal Process., 2011, 21, pp Lippmann, R.: Speech recogniion by machines and humans, Speech Commun., 1997, 22, pp Gu, L., Rose, K.: Percepual harmonic cepsral coefficiens for speech recogniion in noisy environmen. icassp, IEEE, 2001, pp De Cheveigne, A.: Separaion of concurren harmonic sounds: Fundamenal frequency esimaion and a ime-domain cancellaion model of audiory processing, J. Acous. Soc. Am., 1993, 93, pp Nishi, K., Ando, S.: An opimal comb filer for ime-varying harmonics exracion, IEICE ransacions on Fundamenals of Elecronics, Commun. Compu. Sci., 1998, 81, pp Lim, J., Oppenheim, A., Braida, L.: Evaluaion of an adapive comb filering mehod for enhancing speech degraded by whie noise addiion, IEEE rans. Acous., Speech Signal Process., 2003, 26, pp Nehorai, A., Pora, B.: Adapive comb filering for harmonic signal enhancemen, IEEE rans. Acous., Speech Signal Process., 2003, 34, pp Jang, Y., Chicharo, J.: Adapive IIR comb filer for harmonic signal cancellaion, In. J. Elecron., 1993, 75, pp King, B., Alas, L.: Coheren modulaion comb filering for enhancing speech in wind noise. Proc. In. Workshop on Acousics Echo and Noise Conrol, & he Insiuion of Engineering and echnology 2013 IE Signal Process., pp. 1 9

9 26 Schwarz, D., Howe, C., Purves, D.: he saisical srucure of human speech sounds predics musical universals, J. Neurosci., 2003, 23, pp Speech processing, ransmission and qualiy aspecs sq); disribued speech recogniion; exended fron-end feaure exracion algorihm; compression algorihms; back-end speech reconsrucion algorihm, ESI ES v1.1.1, abrikian, J., Dubnov, S., Dickalov, Y.: Maximum a-poseriori probabiliy pich racking in noisy environmens using harmonic model, IEEE rans. Speech Audio Process., 2004, 12, pp raunmuller, H., Eriksson, A.: he frequency range of he voice fundamenal in he speech of male and female aduls, PhD hesis, Manuscrip, Deparmen of Linguisics, Universiy of Sockholm, accessed May ) hp:// hm, ize, I., Marin, D.: Principles of voice producion, Acous. Soc. Am. J., 1998, 104, pp Baken, R., Orlikoff, R.: Clinical measuremen of speech and voice Singular Pub Group, 2000) 32 Barle, R., Sherber, D.: Inroducion o real analysis Wiley New York, 1982), vol. 2 7 Appendix 7.1 Appendix 1. Mahemaical derivaion 1 Mahemaical seps in he inegraion of 5) 7) see equaion a he boom of he page) 7.2 Appendix 2. Mahemaical derivaion 2 Mahemaical derivaion of processing 7) 9) using squeeze heorem. Given by squeeze heorem sin x lim = 1 nb m 0 x see equaion a he boom of he page) B VH 0, s = 0 sin 2mp N s B = N A n sin 2np 0 s + k n ds A B n cos 2p m s 2 0 B + n 0 s + k )) n cos 2p m s B n ) ) 0 s + k n d s [ A n B 1 m nb sin 2p 4p m nb B + n 0 + k )) n s 1 m + nb sin 2p m + nb B s n )] B ) 0 + k n = N = N A n B 2p 1 m nb cos np B + 2 ) nbp 0 + 2k n sin 1 + m + nb cos np )) B 2 ) nbp 0 2k n sin ma n 2 B cos np p m 2 2 n 2 B 2 B + 2 ) npb 0 + 2k n sin = N 0 VH 0, = lim nb m 0 ma n 2 B cos np p m 2 2 n 2 B 2 2 ) npb 0 + 2k n B sin 2 ) npb 0 + 2k n B sin ma = lim n 2 B nb m 0 pm + nb)m nb) cos np = cos 2np ma 0 + k n lim n 2 B nb m 0 pm + nb)m nb) = cos 2np 0 + k n lim = ma n 2np cos 2n 0 + k n nb m 0 nbp sin mp ma n 2 B m nb)p sin pm + nb)m nb) IE Signal Process., pp & he Insiuion of Engineering and echnology 2013