Low-Delay 16 kb/s Wideband Speech Coder with Fast Search Methods

Transcription

1 Low-Delay 16 b/s Wdeband Speech Coder wth Fast Search Methods M. HALIMI M. BENGHERABI A. KADDAI Speech Codng eam Centre de Développement des echnologes Avancées Haouch Oul BP. 17 Baba Hassen Algers ALGERIA Abstract- In ths paper, a low delay 16 b/s wdeband speech coder wth a bufferng delay of 1.5 ms s ntroduced. hs coder s bascally nspred from the G.78 LD-CELP standard for narrowband speech sgnals. Our man goal s to reduce the mplementaton complexty of our wdeband G78 le - coder, whch s manly due to the search of the optmal exctaton (gan-shape) n codeboo. For ths reason an algebrac codeboo s proposed and an exhaustve optmal search and suboptmal full poston and Jont Poston and Ampltude Search technques are mplemented. Objectve and subjectve measures on a large corpus of a testng speech database show that the JPAS (Jont Poston and Ampltude Search) multstage search technque proposed recently can yeld to a very mportant reducton of coder computatonal load, wthout decreasng the perceved qualty of wdeband speech sgnals compared to other search technques. Keywords: - Low Delay CELP, G78, Bacward LPC analyss, Algebrac Codeboo 1- Introducton and overvew of the G. 78 Hgher bandwdths than that of the 300 to 3400 Hz telephone bandwdth result n major subjectve mprovements n speech qualty [1, ]. A bandwdth of 50 to 7000 Hz not only mproves the ntellgblty and naturalness of speech, but adds also a feelng of transparent communcaton. Major applcaton areas of wdeband speech coder nclude Audo and vdeo teleconferencng, hgh-qualty voce transmsson n dgtal AM rado, etc. he qualty reference for wdeband Speech coders has been a subband ADPCM wdeband speech coder snce 1986, commonly now as IU- standard G. 7 at 64, 56 and 48 bt/s [3, 4]. In order to reduce these too hgh bt rates wthout a loss of qualty, wdeband CELP coders, have been developed as an alternatve to ADPCM. he Adaptve Mult-Rate (AMR -Wdeband) codec was standardzed by ESI: 3GPP n December 000, and selected and approved by the IU- n June 001-Januray 00. hs IU- standard wth a frame sze of 0 ms and a loohead of 5 ms s referred to as G.7.. hs standard utlzes the ACELP (Algebrac Code Exctaton Lnear Predcton) technology whch s employed n AMR-Narrowband speech codec as well as IU- G.79 and G.73.1 at 5.3 bt/s. A serous problem wth CELP coders s the codng delay. ypcal encodng/decodng delay of current CELP technology based wdeband audo codecs s more than 40 ms, however n certan bdrectonal and multdrectonal real-tme audo applcatons there s a need for low delay technques. For narrowband sgnals the low delay constrant s realzed by the G.78 [5], whch s a 16 bt/s coder developed by the A& and standardzed n 199. he nput speech to the analyzer s grouped nto small fve-element vectors, and s encoded usng an exhaustve analyss-bysynthess closed-loop search procedure n a traned codeboo. Because of the small frame consdered n the LD-CELP t uses a 50 th order LPC analyss done on the synthess sgnal (bacward adaptve system), and not on the nput sgnal. here s no a long term predcton flter as n the conventonal CELP [6]. he G.78 s very effcent for ncreasng transmsson networs capactes and for multmeda applcatons, but t has a rather a hgh complexty due manly to the search n traned codeboo and the computaton of the 50 th coeffcents of the AR model. A verson of the G.78 codec for wdeband speech at a samplng rate of 3 Hz has already been proposed [7]. In ths paper t s suggested that the end to end delay should be less than 5 ms. A frame length of 1.5 ms whch corresponds to 0 samples of wdeband speech at 16 Hz samplng rate s consdered. Our coder s actually a modfcaton of the G.78. A major modfcaton to the conventonal algorthm n the current paper s that

2 the traned codeboo n [5], or the algebrac ternary codeboo n [7] are replaced by 15 bts bnary algebrac shape codeboo and 5 bts gan codeboo. Snce the exhaustve search n a 15 bt algebrac codeboo s too expensve, suboptmal approaches are appled as poston exhaustve search (bacward sgn) and jont poston and ampltude search (JPAS). A man feature of our codec s that an exctaton of dmenson 0 s ndexed wth 0 bts, so we have a quantzaton rate of 1bt/dmenson whereas n prevous wors they had a rate of bts/dmenson [5, 7]. he organzaton of ths paper s as follows. In secton we gve a descrpton of the proposed lowdelay 16 b/s wdeband speech coder and essence of codeboo search. In secton 3, we ntroduce a 15 bts algebrac codeboo and fast search technques. Secton 4 s dedcated to smulaton and Coder performance under dfferent search technques. Fnally, a concluson s gven.. Descrpton of the proposed coder he wdeband coder presented n ths paper s a modfcaton of the low-delay speech codng algorthm G. 78. In a prevous wor [7] a hgh qualty 3 bps wdeband LD-CELP was proposed, but the computatonal load of the encoder maes the real tme mplementaton mpractcal on a sngle DSP3c. nowng that ths computatonal load s manly due to the search of the optmal exctaton n the traned codeboo, Murga et al [7] have proposed to use a ternary exctaton, ths yeld to a computatonal gan of 8% of the total codng tme. he major modfcaton n our coder s that the frame length s ncremented to 0 samples; ths corresponds to an algorthmc delay of 1.5 ms at a samplng frequency of 16 Hz. 0 bts are used to ndex the exctaton codeboo whch results n a rate of 16 b/s..1 Essence of codeboo search In order to reduce the complexty search through 0- bt codeboo a "gan-shape" codeboo structure was adopted. he 0-bt codeboo s decomposed on two codeboos: the frst one wth 15 bts for shape quantzaton and the other one wth 5 bts for the gan quantzaton. he search procedure conssts n scrutnzng the shape codeboo and dentfyng the best ndex of gan for every shape codeword C After the dentfcaton of the best ndces and, these last ones are concatenated to form a word of 0 bts. he codeboo search s made accordng to the mnmzaton crtera of the followng MSE dstorton: D = x x, (1) x: s the target speech vector; and x s gven by x σ g (), = H C where H s a lower trangular convoluton matrx constructed from the mpulse response of the weghted synthess flter, g and C are, respectvely, the th level n the gan codeboo and the th codeword n the shape codeboo and σ s the bacward adapted exctaton gan. Usng () equaton (1) can be wrtten as: D g = σ x HC (3) D where x s the normalzed target vector. If = 0 s g taen, equaton (3) gves: Replacng (4) n (3) we get x HC g opt = (4) C H HC ( x HC ) D = σ x x (5) CH HC he mnmzaton of D reduce to the maxmzaton of the followng term ( x y ) = (6) α where y = HC s the response of the flter H to the nput vector C and α =y y s the energy of the fltered vector. can be wrtten as ( d C ) = (7) α where d = H x s the bacward fltered normalzedtarget vector. he computaton of α remans the man problem n CELP coders wth random codeboos and hgh dmensonal exctatons. But, the use of algebrac codeboos wth very effcent suboptmal search technques solves ths problem [9].

3 3. he algebrac shape codeboo desgn An nnovaton vector of dmenson 0 contans 5 nonzero pulses. he fve pulses can assume the ampltudes and postons gven n table I. he ampltude of the 5 pulses can be ether 1 or 1. Each pulse poston s encoded wth bts, and each pulse ampltude s encoded wth 1 bt, resultng n a 15 bts codeboo. he followng table summarzes the structure of dctonary. able 1: Structure of algebrac codeboo rac Pulse Ampltude Postons Bts 1 0 ±1 1, 6, 11, ±1, 7, 1, ±1 3, 8, 13, ±1 4, 9, 14, ±1 5, 10, 15,0 1+ he codewords n ths codeboo can be represented by ponts unformly dstrbuted over the surface of a hyper sphere of 0-dmensonal space. Any codeword C from ths structured dctonary s gven by 5 = 1 ( m ) C = bδ (8) where δ () 0 = 1, δ () = 0 for 0 and b s the pulse ampltude. he pulse postons are gven by: ( ) = 1, L,5 m j = + 5 j (9) j = 0, L,3 he term n equaton (7) can be wrtten as ( d C ) = (10) CΦC where Φ=H H s a matrx contanng the autocorrelatons of the mpulse response of the weghted synthess flter H. As the vector C contans only 5 nonzero pulses, equaton (10) can be wrtten as ( m ) = 1 = 5 (11) Φ b 4 5 ( m, m ) + bb jφ( m, m j ) = 1 = 1 j = d Usng an exhaustve search a total of 15 combnatons are performed n 5 nested loops, each loop corresponds to one pulse poston. However, ths approach s not practcal n real tme coders, so we must apply suboptmal but faster search technques. If the ampltudes of pulses are settng a pror to the sgns of bacward fltered vector d(n), ths choce decreases the number of combnatons to 10 (104). hs suboptmal technque s called poston exhaustve search (bacward sgn). Another nown suboptmal search technque called the focused search can be used [9]. In ths case a very small subset of the codeboo s searched. However, n our wor we have used the JPAS (Jont Poston and Ampltude Search) ntroduced recently by Ramrez and Geren [10] as an alternatve of focused search brngng about a reducton n complexty of one thrd n comparson to the focused search. Applyng ths powerful technque n searchng the optmal exctaton n our codeboo reduces the number of combnaton to 100 (0 for the frst pulse + 3 for the second pulse + 4 for the thrd + 16 for the fourth + 8 for the last). So a very sgnfcant reducton n the number of combnaton s obtaned wth respect to poston exhaustve search and the standard exhaustve search where ths number s 104 and 3768 respectvely. Fnally, we should menton that n order to determne the optmal gan quantzer, we frst create a database consttuted by the optmal gans obtaned durng the encodng of a speech tranng sequence. We then apply the LBG algorthm [11] to have a 5-bts codeboo (sgn bt ncluded). 4. Smulaton and coder performance under dfferent search technques In order to evaluate the performance of the proposed coder and to compare the prevously mentoned search technques, the source test materal conssted of the IMI database [1]. From ths database, we selected eght male and female speech sgnals from dfferent dalect regons. Each speech sgnal contans 30 sentences spoen by seven speaers. he speech sgnals characterstcs are lsted n the table. he nput speech sampled at 16 Hz s fltered by the sendng flter p.341, and the speech level s adjusted at -6dB.o evaluate the performance of both methods, t s necessary to use the two man test categores: objectve measures (SNR and segmental SNR) and subjectve ones carred out accordng to IU recommendatons P.800/P.830. he results of these tests are averaged to gve mean opnon scores (MOS). Such tests are often tme consumng and costly as they requre proper tranng of lsteners. Some objectve measures that have been standardzed by IU- le

4 the PESQ (Perceptual Evaluaton of Speech Qualty) technque [13] have been shown to provde a score that s strongly correlated to the subjectve ratng obtaned by MOS. Although ths measure has not been fully characterzed for wdeband speech tll now, we have used t to estmate the MOS of the proposed coder. It should be mentoned that n the calculaton of the segmental SNR (SEGSNR) wth frame length equal to 56 speech samples (16 ms), we elmnate the slence perods from the calculaton. here are stated as speech ntervals wth power below 40 db. able. Sentences sets characterstcs used to evaluate he objectve and subjectve performance. Sentences he dalect regon Speaers No. of Samples 1 DR1 Male DR Female DR3 Male DR4 Female DR5 Male DR6 Female DR7 Male DR8 Female dsplay the computer tme executon (PC wth 866 MHz Pentum Processor under Wndows XP operatng system) requred by each search method n the fxed exctaton dctonary. Smulatons results show that usng JPAS search scheme the codeboo search tme s approxmately 6 % of the tme requred by the full poston search and represents only 0.3% of the exhaustve search technque. able 3. he objectve performance of three ACELP searches for sentences sets 1-8. SNR/Segmental SNR Exhaustve BS JPAS Search Search Search / / / / / / / / / / / / / / / / / / / / / / / /11.37 able 4.he subjectve performance of three ACELP searches for sentences sets 1-8. able 3 presents the objectve performance n terms of SNR and SEGSNR of the three search technques where the JPAS search yelds to dmnsh of 1.5- db n SNR and SEGSNR as compared to exhaustve search, but only a slght dmnsh less than 0.5 db compared to the suboptmal full poston search technque (BS). able 4 presents the estmated MOS under dfferent search technques for the speech sgnals lsted above. he hghest PESQ results of the exhaustve Search show that our coder nspred from the LD CELP G.78 but at a quantzaton rate of 1 bt/dmenson has only a qualty slghtly better than far, so t does not acheve the needed good qualty needed n practce. However, an mportant remar s that the PESQ measures show that the perceptual performance dfference s small between the JPAS and the exhaustve search technques and very small and nsgnfcant between the JPAS and the full poston search technques. Fnally, for the evaluaton and the comparson of the dfferent search technques, we PESQ Exhaustve BS JPAS Search Search Search Concluson In ths paper, a wdeband G78 le - coder wth an algorthmc delay of 1.5 ms s presented. he man feature of ths coder s the use of an algebrac codeboo. Subjectve evaluaton of the proposed coder show that our coder nspred from the LD CELP G.78 but at a quantzaton rate of 1 bt/dmenson has

5 only a qualty slghtly better than far, so t does not acheve the needed good qualty needed n practce. However, an mportant remar s that the perceptual performance dfference s small between the JPAS and the exhaustve search technques and very small and nsgnfcant between the JPAS and the full poston search technques. Fnally, we beleve that dong some modfcatons on the perceptual flter may mprove the performance of our coder. References: [1] P. Noll, Wdeband Speech and Audo Codng, IEEE Communcaton Magazne, vol. 6, Nov. 1993, pp [] N. Jayant and Y. Shoham, Codng of wdeband speech, Speech Communcaton, vol. 11,199, pp [3] CCI recommendaton G.7, 7 Hz audo codng wthn 64 t/s, n Blue Boo Fasccle III, Oct [4] P. Mermelsten, G. 7, A New CCI Codng Standard for Dgtal ransmsson of Wdeband Audo Sgnals, IEEE Communcatons Magazne, Jan. 1988, pp [5] J.H. Chen, Hgh Qualty 16 b/s Speech Codng wth a One Way Delay less than ms, Proc. IEEE Int. Conf. on Acoust. Speech Sgnal Processng, Apr. 1990, pp [6] M. R. Schroeder, B. Atal, Code Excted Lnear Predcton (CELP): Hgh Qualty Speech at Very Low Bt Rates, Proc. IEEE Int. Conf. on Acoust. Speech Sgnal Processng, Apr. 1985, pp [7] C. Murga, G. Feng, A. leguyader, and C.Qunqus, Very low-delay CELP and hgh qualty of 0 hz 15 hz speech sgnals at 64 bt/s, n Proc. IEEE Int. Conf. on Spoen Lang Processng [8] O. Gottesman, Y. Shoham, Real-me Implementaton of Hgh-Qualty 3 KBPS Wdeband LD-CELP Coder, he European Conference on Speech Communcaton and echnology (EUROSPEECH, 1993). [9] C. Laflamme, J-P. Adoul, R. Salam,S. Morssette, P. Mablleau, 16 bps Wdeband Speech Codng echnque Based on Algebrac CELP, Proc. IEEE Int. Conf. on Acoust. Speech Sgnal Processng, oronto, CA, May 1991, pp [10] M. A. Ramrez, M. Geren, Jont Poston and Ampltude Search of Algebrac Multpulses, IEEE rans. Speech Audo Processng, vol. 8, no. 5, Sept. 000, pp [11] Lnde, Buzo and Gray, An Algorthm for Vector Quantzaton Desgn, IEEE rans. Comm., Jan [1] NIS, he DARPA IMI Acoustc-Phonetc Contnuous Speech Corpus, Oct [13] IU-, Recommendaton P.86 Perceptual Evaluaton of Speech Qualty (PESQ), an Objectve Method for End-to-End Speech Qualty Assessment of Narrow-band elephone Networs and Speech Codecs-, Feb. 001.