ICASSP2019音声&音響論文読み会 「ICASSP2019における音声認識の最新技術動向」#yjtc #icassp2019jp

スライド概要

ICASSP2019音声&音響論文読み会 「ICASSP2019における音声認識の最新技術動向」 @ DeNAオフィス
by ヤフー 三宅 純平

profile-image

Yahoo!デベロッパーネットワーク

@ydnjp

作者について:

エンジニア・デザイナー向けのヤフー公式アカウント。イベント/登壇情報/ブログ記事など、ヤフーの技術・デザインに関わる情報を発信します。

スライド一覧
シェア
埋め込む»CMSなどでJSが使えない場合

公開日

2019-06-07 16:27:00

各ページのテキスト

1. ICASSP2019 P A 9 1 ) 6 72019 ) 1 6 ) 76 7 92 ( 2019 7 . 0 6 2 C 8 P I

2. l V v n (9 2 r s s j s Y @ 13 3 3 @ @3 ) Y Ry k v _kt_k g e 1 0. Ep O m A A @C 4 a Jhd Eo s I hd i J hd

3. • C A • C 0 • 9 • • ( 1 0 ( 0 ) 12 3( 1 3 22 1 0 . . 9.

4. e ar K v ihos J y ( 12 ( .2 ). . ( K . 44 12 9 090 0 v Y O pt CA l g v n n d R

5. S R ? R y 21 s A a eTN ld O w A9 C 5 S 5 5 A h A5C gY ) 9 C A R oKS id 3 5C C 5C A 1 rt 0 . JJ ʼk A bu kav p nu VW

6. A • • .. Y RN • ) HA 9 C ) H A T )2) (CC .D C 1CH C A C A LJM -. ( C C ) C S .D C 9D (10 ) , ) A A C ( 09 C 0 C A . 2A 9 C A C CD (10 1 C 0 9 2

7. ( 1 0 ( 0 ) ( 1 22 1 07 .7. 9.

8. Ys K h r J p ( 8 12 ( .2 ). . ( 8 8. R 12 9 K t 0908 0 g on CA i Od el a hel

9. R Y JA C JA ) 2019 ) 1 ) 92 ( 2019 .

10. ( 6 H C ) ) A J ( C + 1 0 0 ) 1 22 1 0 0 6 . . 9. 3 6

11. T D) , M *) HR y M pq D M D M , E I , ( * 9 i 19E9C , E Y 9I C b ans c dNege c , t ( / IH H J ( ( _ 9CA L8L c YSa_ c ba F( o - F ( / Yv C nsY lr / k C D, t A - A lrh ( R T MHEL N nsc t -C I -C 2. 0

12. e ka to M M . 2 . . a . . M A e R Yd N 9 12 .2 (. . 9 9. N 12 0 09 0 J ) CJL

13. J A ( 1 0 ( 0 ) 3( 1 3 22 1 0 . . 9. J C C

14. (4 1 0 ( 044 ) (4 4 14 22 1 0 . . 9.

15. T C) N N C , ( * ISa qr N0 CA n _ s Y / Nyl v” E H C, -EFLHA J - ( 8 EE 2 F -EHFEH JAE ,CC A acY bot d e - sig C R HN d , _ ( / A JI I H d acb E H - C ( ( ( / h k n 55 ( p Y M9M d b R M E T L D otd Y . JE . 1 )

16. -- + 530 D R N 5 9 a xs Y y 0 8 65 H N 5 9 g : J C : 45 15A5 A C A F C A F C A | 9 M+ 5 . 3: 3 . k . 1 |v C A F C C 38C8 8 ) -- -- el (( tv lr v D dir s p f b (+F2 Y ( o ih) a a l n( ) ) ( -dhr

17. D C 199a pjnr vi Ysn 4 A P AO DG D - d 4 L FOA f pjnr Ysn r ML DT 1DNDLCD A A R AH P 0MN OHFG 0 + AGMM ANAL 0MONMOA HML /JJ <HFG P <DPDOSDC A H Yl 2 > a u GMLD DTH ML u A R IANA9 fl 07 d 9 FOA DH OHNGMLD u h n :>/a g7A H D 2ODD 5 h c h c u ytY mxs u -P. ytY ytY mxs mxs - P. Yo u ytY () mxs -DNP. ) Yo ( aeW W

18. 81 J • C C C J A C J C Kim, Chanwoo, et al. "Efficient implementation of the room simulator for training deep neural network acoustic models." arXiv preprint arXiv:1712.03439 (2017). 9 0. ( 2 0 2 11 )0. )

19. 9 0. ( 2 0 2 11 )0. )

20. T D) , ( M *) HR y t M pq D M D M D, , E I , ( * 9 19E9C , E / ” 9I C c , Y t ( / IH H J ( ( / Yv C lr i k C A ( ans c dNege - A lrh b R c ba - _ 9CA L8L c YSa_ F( o T MHEL N F nsc t -C I -C 2. 0 (

21. E J rb nt x chl ph Y • 0 E) • E a • - - A 9:C ( : . 21 A A C chl ) s 1 9:CB 1 B A B d oil rba N g E x // A : / + BC m S S E - E //s NY NY s vh RSP | bvyea RYP

22. o a A A d J S R bk A iYh d p L r s 2 g el n t gci (0 wygci 22(02 2 (200 222 (02 v MTgci _f gi ) A C k 0- v MTgci , 0 A9 C 60 1 0 ,6 0. A A C 0 : ,6 .6, ( . 9 CB . B A ., 6, 066 2 02 0 0 0 0 .0 , , A 0. 0 0 ,6 0 A ""

23. 6GN DJ A DJC AJC LDII 62 EID BGBD DAMB /6088 3 a - ANH 1 88 ry nloWv w aSs d V e gT Y S v S t n h / ( , A %6 DNB MGJE /- 9( c v .G CGMDB G JAI 6 kpT d l iod .: 6 6. 2 . 6 2: . 2:1 .:1 2 0 :A2 . 6 :. 220 20 :6 6 : ,, / LRMGE d EID JG CGMDB G JAI 6 - DJ G J U ) 2 - ALAJ / ML MA G J -II GE N DNDMPDC " :2 . :2 . 2 A 0. .C ()

24. / AS 1 1 f R BT JG A C 9 h 999 cb cpv u r h a YiSc0 y c 1 1f s d Rl di R ta H T l W cR o nj l“ A g Y o” “ e cpv CDOM pv u ro u rd o 79 9 9: " 5 A A" 1 A 1 ( . 89 39 9 1 9A 0CC 4A H4 H 9C 2 : A 1 1 9 A 9 A J - ( , / ( H ( ) AA (

25. ) ) o = CY ap • g le i d lR T T n • CY ap hR T T J n A = K. Rao, H. Sak, and R. Prabhavalkar, Exploring architectures, data and units for streaming end-to-end speech recognition with rnn-transducer, in Proc. ASRU, 2017, pp. 193–199. ) 12 ) .2 . .5 ) . 5 ( 12 9 090 0 T

26. (0- 1 (0)0 2 / - EC : A ED E =G =D = JE =G =D = E =B 2EE B= EA = =: J: A Shr E A AD:B+ ( E mP.: =BAD=T qp cig efR 0 5 ED= o C A M 1 1 n c dY7 A J, ad ,ql s . . . 7. , - 220 ,20 E M A J 9: EE : :D E 2 ." "C E :JAED -BB 6A . 2 220 " J 6= = = -2 A2 02 " -2 A2 02 E == 6= E DAJAED 12 )

27. )/ ( -H 7HH C 2 9 oidq g a C J HC C J HC " (, CDL = H 2 - 8 M H7H C H 7 H D 9 9C= H C H 7 D Rhp 0C 0 H 7 7HH H C HH H C 7 ACCH =r n W W lf Y b 2 7 ) , ,11 -00 7S Psc me "2 . , 2 ( 7 CC .7D7 10 )/ C DC 7H C 2 2 2 2 02 = H J )0 220 20 0 " . 220 A 1. 2 2 02 - 02 2 2 02 8 12 " , )..- , 9 ( HC 9 AC N

28. ( 1 08 ( 0 ) ( 81 22 1 08 . . 9.

29. ) rC Yht J lg O oa 2019 1 ( 92 2019 ei JAsp . l R K n d lgd

30. n a s A2 ( 9 23 ( . 3 i r : Rs Kd pWS gvWS o 5 . ) ( 9 9 le 23 :: 5 1 19 10 0 . '5. J Cth s Y r 5: 5 0 '5 '5: 0 A y

31. C ) • • • in d K R in R d 1(0 v LP 2 y cK LP oh JR d • c N v RV u g R d ( A YeS cK A V s t l A: 0 A5 2 CC: 1 A 5 " 5- A . : 9A C 20 l pJr W u I, ", A ).a RV WLR d 2G K C , "3 9 g LP K • 9A ( D • D B - ( ( A ) A 5 2 4 A A C A: C , C: A5 CC: 2 : 4 5 C A A 4 H C A 20 ) , 221 (

32. ** * • 79 ( 79 " .C E . ( : 9 A * , , * -* c N -AA 1 YN N CER fe ialpo hn g - 9 0HEJ M 9 CC 29 9 .CE CE9 C 1 ) ( 9A EI 3 JCE C CE CC A 9 9 H * :C 1 d K H9A ( * S. : E :C g C )(

33. )( ( 9 23 ( . 3 ( 9 9 23 1 19 10

34. • , , 21 332 dh Sv 21 • l t 2E C • c i NTPS • fen J rp , , s , , T2 9 , , Y T 21 CC C a g w NR 2 A C )4 . A A C 1 C 1 A 0A o G, C , G y (

35. ba h. : EA . : ien : A EA , E , , E EA M iY :A , J, 0 : p Ogl STco J , 5 9AEACA : p gl r -AB C E - ) AA B -ACBAC E A 3 E 3 C : L OMR d O12 (

36. 776 EC • 6 4 6IMUI 3SN g • 64 2- MITE IR N 1AKI N MIA 8N d u 8 s np 55 MGEKE MIRE ad W , k 6 4 4 6 1 6 4 5 6 6 2- 7 :7 7 64 . 2 HIGEJI 7A IRA ER AK 7 65 DH 21 5 2 RARE . A A pk 2 7 7 1NLLSMICARINM CIEMCE 8ABN ARN IE 2 AIMARH ;NM 6 s 2 6A AM 64 56 7 EI 3NNGKE s 2 MC .5 64 MIRE 6 4 4 5 -2 7 2 26 67 2 AJAAJI 4N I ER AK 9IR SBI HI 2KECR IC ;E EA CH 8ABN ARN IE 75 2 2- 5 2 F IGHR 1 AG % - AHNN 6A AM 1N N ARINM KK ;IGHR ;E E TE 7 2 2 i 2 -2 7 64 n . 92;8 ACRN NM BA E g ;kEMCN E s 2MCN E 2 s 2 o ; 2MCN E 0ACJ N s dm1 CKE CNM I REMC KN %%HNS t x oW( %HNS kt t xs ef ) 1N RARE -2 7 SM AI k h v o ;h g u y d SRNEMCN E s AI xg dEMCN E h SRNEMCN E kEMCN E kbrcrk8N 2;g % Y o , %Ylg • .5 s 2; (

37. . ED B 5A CC 5 • SFCI -I 1 1 MCS 2RPJ FJ 9 MH 1KKDIC 3JA CJAK CM CAK CM uyporh) 5 • - : 5 5-5 1 1 e - - 5 1 JF C 0 I RCM a V % p v z ” “ eik5 2CR FMFU U IIPAFJ CNUe “ tn j5 u s c)%Y . % , KK L J .KMLKM FKJ -II FD N CNCM C . :9 CN JF C 5 FGL -1 W - 1 5-1 - 9 .5 KKG R CKJ 6CK 6FP 2CJMR 7 NKJ -LLIC .KLRMFD -1 FHIK - 1 5 - CN a ag d k glk uy 6 FAC (

38. • • • 9 -33 /22 /H D e d hb .A .A e d hb , , A A /H D , , A .A .A Y i hb - 0 C1 .A -33 /22 , CHD , 89 09C9A , DC D9 A E E D Jg 33 1 hb .A e A 2 d a KS S WT N MR ()

39. ) ( 9 23 ( . 3 ( 9 9 23 1 19 10