SlideShare a Scribd company logo
1 of 35
Download to read offline
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
n G
n
n
n -
n A C
2
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
n a
/ []eb A D a C
/ D 4 1 /0 , 1 6
3
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
n I N
)
n A
B B
(
(
(
(
(
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
5
3 0 2 . /3 0 7 3 1 . 0 0 7 3
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
6
3 0 2 . /3 0 7 3 1 . 0 0 7 3
:
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
n
A
B A
7
B
( A )
( ; F0)
( ; bap)
B
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
8
(F0 )
bap
•
→ Vocoder
•
• STRAIGHT [Kawahara+; ’99]
• WORLD [Morise+; ’16]
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
Vocoder
9
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
vocoder
10
F0bap
F0bap
F0bap
1 frame
frame
Frame
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
[Abe+; ’90][stylianou+; ’98]
n
A B
11
F0bap
F0bap
GMM DNN
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
[Abe+; ’90][stylianou+; ’98]
n
B
12
AF0bap
AF0bap
GMM DNN
• F0, bap
→ A
• F0 bap
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
13
Parallel-data
B
A
frame
A
B
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
o s e P P
• 6C C E C AA 6-- ( y
• 6-- K K E 2; ] p e
d r aKvt 6-- (
3 E AA A ; G
• g K V P[P PN g kO
• E AA A ; G P Po - A 1 70 C + )8
i h ced
• nQc i h ʻ] d l ]SP
• nQc 6 6 . 7 ; 2CE; + )8
14
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
ig l hV vNVQ] V
• 3 7 7 E C 7;6 Nk
16 6 6 6Nyo
• Ns V PV cK
• A6 6 6 6 V i + 7 - .6 G
n a N []
• n a r pdNʻ O
• t V e [n 32 3 , E6 0 G
15
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
Voice Conversion Challenge 2016
n
n 7
7
7
n 5 5
n 01
16
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
Results of listening tests in VCC 2016
17
cf. http://vc-challenge.org/vcc2016/summary.html
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
ig l hV vNVQ] V
• 3 7 7 E C 7;6 Nk
16 6 6 6Nyo
• Ns V PV cK
• A6 6 6 6 V i + 7 - .6 G
n a N []
• n a r pdNʻ O
• t V e [n 32 3 , E6 0 G
18
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
CycleGAN [Zhu+; ’17]
n
19
cf. https://junyanz.github.io/CycleGAN/
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
CycleGAN [Zhu+; ’17]
n
n
Forward-inverse mapping Inverse-forward mapping
GX→Y GY→X G L real/fake loss
[Kaneko+; ‘17]
M
mapping loss
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
CycleGAN [Zhu+; ’17]
n
n
21
Forward-inverse mapping Inverse-forward mapping
GX→Y adversarial loss
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
CycleGAN [Zhu+; ’17]
n
n
22
Forward-inverse mapping Inverse-forward mapping
= "#~%&'(' # log,- . + "0~%&'(' 0 log 1 − ,- 34→- .
GY→X adversarial loss
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
CycleGAN [Zhu+; ’17]
n
n
23
Forward-inverse mapping Inverse-forward mapping
L1loss
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
CycleGAN [Zhu+; ’17]
n
n
24
Forward-inverse mapping Inverse-forward mapping
λcyc 10.0
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
CycleGAN parallel-data-free [Kaneko+, ’17]
n NG NG
n
C
25
CycleGAN
copy
A
A
A
A
A
t1 t2 tTbap
bap
bap
bap
bap
F0
F0
F0
F0
F0
bap
bap
bap
bap
bap
F0
F0
F0
F0
F0
A
A
A
A
A
t1 t2 tT
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
VC
. 1 1 2 r
• d c U R l U
• c U
• t pt t G l em
- 1 1 a ) ( A . (1 1 ( l
• y sv cG X
• g I ni L UI o
26
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
Network architecture
n . / - . /
n :
. / / . / / / .
27
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
ig l hV vNVQ] V
• 3 7 7 E C 7;6 Nk
16 6 6 6Nyo
• Ns V PV cK
• A6 6 6 6 V i + 7 - .6 G
n a N []
• n a r pdNʻ O
• t V e [n 32 3 , E6 0 G
28
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
Variational autoencoder (VAE) [Hinton+; '06]
n
z
29
x
Encoder
qθ(z|X)
Decoder
pθ(X|z)
z
!"
# $; 0, 1
Input feature Generated feature
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
VQ-VAE [van den Oord+, ‘17]
n
- -( ) E V A
n
30
x
Encoder
p(ze(x)|x)
Decoder
p(x|zq(x))
ze(x)
!"
A
A
e1 e2 e3 eK
zq(x)
x LQ loss VQ loss Encoder loss
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
VQ-VAE
n [van den Oord+; ’16]
t v o G a r
• N x h G d
• λ W l lg d
• l m r e
31
! " # = %
&'(
)
* +&|+&-),+&-)/0, ⋯ +&-0, #
λ : d lg c d
, " = +(, +0, ⋯ +&-0
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
VQ-VAE [van den Oord+, ’17]
n
32
Encoder WaveNet
ze(x)
e1 e2 e3 eK
zq(x)
id
• zq(x) id
• ze(x) zq(x) id
• zq(x)
( )
https://avdnoord.github.io/homepage/vqvae/
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
VQ-VAE [van den Oord+; ‘17]
n
33
Encoder WaveNet
ze(x)
e1 e2 e3 eK
zq(x)
id
cf. https://www.slideshare.net/YukiSaito8/saito18sp03
• zq(x) id
• ze(x) zq(x) id
• zq(x)
( )
https://avdnoord.github.io/homepage/vqvae/
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
n A
n C
-
34
Copyright © DeNA Co.,Ltd. All Rights Reserved.
Strictly confidential
H9J J JZJ d.-I 6 9J J JZJ 7 J[]MJ 9J[][N JVM 0 MN 1 N NRPVN e N[ Z]L ]ZRVP [XNNL ZNXZN[NV J RWV[ ][RVP J
XR L JMJX R N RUN OZNY]NVLa [UWW RVP JVM JV RV[ JV JVNW][ OZNY]NVLa KJ[NM 4 N ZJL RWV W[[RKTN ZWTN WO J ZNX R R N
[ Z]L ]ZN RV [W]VM[ f XNNL 1WUU]VRLJ RWV , XX -, , ...
H WZR[N c +I WZR[N 4 FWSWUWZR JVM 9 bJ J eD :2 J WLWMNZ KJ[NM RP Y]JTR a [XNNL [aV N[R[ [a[ NU OWZ
ZNJT RUN JXXTRLJ RWV[ f 73713 ZJV[JL RWV[ WV RVOWZUJ RWV JVM [a[ NU[ WT 3.. 2 VW , XX -,, --) +
H0KN . I 0KN JSJU]ZJ 9 RSJVW JVM 6 9] JKJZJ eCWRLN LWV NZ[RWV ZW]P NL WZ Y]JV RbJ RWV f
, ,+
H[ aTRJVW] c.-I F aTRJVW] 1JXXh JVM 3 W]TRVN[ e1WV RV]W][ XZWKJKRTR[ RL ZJV[OWZU OWZ WRLN LWV NZ[RWV f
( ) ..-
H9JVNSW ,I A 9JVNSW JVM 6 9JUNWSJ f JZJTTNT 2J J 4ZNN CWRLN 1WV NZ[RWV [RVP 1aLTN 1WV[R[ NV 0M NZ[JZRJT
N WZS[ f JZER ,
HG ] c ,I 8 F G ] A JZS 7[WTJ JVM 0 0 3OZW[ e VXJRZNM RUJPN W RUJPN ZJV[TJ RWV ][RVP LaLTN LWV[R[ NV
JM NZ[JZRJT VN WZS[ f
H6RV WV c +I 5 3 6RV WV JVM JTJS ] MRVW e NM]LRVP N MRUNV[RWVJTR a WO MJ J R VN]ZJT
VN WZS[ f ,-+ ) , +
H JV MNV WZM c ,I 0 JV MNV WZM JVM CRVaJT[ e N]ZJT MR[LZN N ZNXZN[NV J RWV TNJZVRVP f 7V
XX +( . +( - ,
H JV MNV WZM d +I 0 JV 2NV WZM 2RNTNUJV 6 GNV 9 RUWVaJV CRVaJT[ 0 5ZJ N[ JVM 9 9J ]SL]WPT]
eDJ NVN 0 PNVNZJ R N UWMNT OWZ ZJ J]MRW f
35

More Related Content

What's hot

分布あるいはモーメント間距離最小化に基づく統計的音声合成
分布あるいはモーメント間距離最小化に基づく統計的音声合成分布あるいはモーメント間距離最小化に基づく統計的音声合成
分布あるいはモーメント間距離最小化に基づく統計的音声合成Shinnosuke Takamichi
 
深層生成モデルに基づく音声合成技術
深層生成モデルに基づく音声合成技術深層生成モデルに基づく音声合成技術
深層生成モデルに基づく音声合成技術NU_I_TODALAB
 
自然言語処理のためのDeep Learning
自然言語処理のためのDeep Learning自然言語処理のためのDeep Learning
自然言語処理のためのDeep LearningYuta Kikuchi
 
Nakai22sp03 presentation
Nakai22sp03 presentationNakai22sp03 presentation
Nakai22sp03 presentationYuki Saito
 
深層学習を利用した音声強調
深層学習を利用した音声強調深層学習を利用した音声強調
深層学習を利用した音声強調Yuma Koizumi
 
統計的音声合成変換と近年の発展
統計的音声合成変換と近年の発展統計的音声合成変換と近年の発展
統計的音声合成変換と近年の発展Shinnosuke Takamichi
 
LSTM (Long short-term memory) 概要
LSTM (Long short-term memory) 概要LSTM (Long short-term memory) 概要
LSTM (Long short-term memory) 概要Kenji Urai
 
Non-autoregressive text generation
Non-autoregressive text generationNon-autoregressive text generation
Non-autoregressive text generationnlab_utokyo
 
雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習
雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習
雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習Shinnosuke Takamichi
 
音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3
音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3
音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3Naoya Takahashi
 
[DL輪読会]Focal Loss for Dense Object Detection
[DL輪読会]Focal Loss for Dense Object Detection[DL輪読会]Focal Loss for Dense Object Detection
[DL輪読会]Focal Loss for Dense Object DetectionDeep Learning JP
 
文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)
文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)
文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)Shirou Maruyama
 
実環境音響信号処理における収音技術
実環境音響信号処理における収音技術実環境音響信号処理における収音技術
実環境音響信号処理における収音技術Yuma Koizumi
 
Numpy scipyで独立成分分析
Numpy scipyで独立成分分析Numpy scipyで独立成分分析
Numpy scipyで独立成分分析Shintaro Fukushima
 
Attentionの基礎からTransformerの入門まで
Attentionの基礎からTransformerの入門までAttentionの基礎からTransformerの入門まで
Attentionの基礎からTransformerの入門までAGIRobots
 
[DL輪読会]GANSynth: Adversarial Neural Audio Synthesis
[DL輪読会]GANSynth: Adversarial Neural Audio Synthesis[DL輪読会]GANSynth: Adversarial Neural Audio Synthesis
[DL輪読会]GANSynth: Adversarial Neural Audio SynthesisDeep Learning JP
 
異常音検知に対する深層学習適用事例
異常音検知に対する深層学習適用事例異常音検知に対する深層学習適用事例
異常音検知に対する深層学習適用事例NU_I_TODALAB
 
Onoma-to-wave: オノマトペを利用した環境音合成手法の提案
Onoma-to-wave: オノマトペを利用した環境音合成手法の提案Onoma-to-wave: オノマトペを利用した環境音合成手法の提案
Onoma-to-wave: オノマトペを利用した環境音合成手法の提案Keisuke Imoto
 
DAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也hare
DAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也hareDAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也hare
DAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也harePreferred Networks
 

What's hot (20)

分布あるいはモーメント間距離最小化に基づく統計的音声合成
分布あるいはモーメント間距離最小化に基づく統計的音声合成分布あるいはモーメント間距離最小化に基づく統計的音声合成
分布あるいはモーメント間距離最小化に基づく統計的音声合成
 
深層生成モデルに基づく音声合成技術
深層生成モデルに基づく音声合成技術深層生成モデルに基づく音声合成技術
深層生成モデルに基づく音声合成技術
 
自然言語処理のためのDeep Learning
自然言語処理のためのDeep Learning自然言語処理のためのDeep Learning
自然言語処理のためのDeep Learning
 
Nakai22sp03 presentation
Nakai22sp03 presentationNakai22sp03 presentation
Nakai22sp03 presentation
 
研究効率化Tips Ver.2
研究効率化Tips Ver.2研究効率化Tips Ver.2
研究効率化Tips Ver.2
 
深層学習を利用した音声強調
深層学習を利用した音声強調深層学習を利用した音声強調
深層学習を利用した音声強調
 
統計的音声合成変換と近年の発展
統計的音声合成変換と近年の発展統計的音声合成変換と近年の発展
統計的音声合成変換と近年の発展
 
LSTM (Long short-term memory) 概要
LSTM (Long short-term memory) 概要LSTM (Long short-term memory) 概要
LSTM (Long short-term memory) 概要
 
Non-autoregressive text generation
Non-autoregressive text generationNon-autoregressive text generation
Non-autoregressive text generation
 
雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習
雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習
雑音環境下音声を用いた音声合成のための雑音生成モデルの敵対的学習
 
音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3
音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3
音源分離 ~DNN音源分離の基礎から最新技術まで~ Tokyo bishbash #3
 
[DL輪読会]Focal Loss for Dense Object Detection
[DL輪読会]Focal Loss for Dense Object Detection[DL輪読会]Focal Loss for Dense Object Detection
[DL輪読会]Focal Loss for Dense Object Detection
 
文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)
文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)
文法圧縮入門:超高速テキスト処理のためのデータ圧縮(NLP2014チュートリアル)
 
実環境音響信号処理における収音技術
実環境音響信号処理における収音技術実環境音響信号処理における収音技術
実環境音響信号処理における収音技術
 
Numpy scipyで独立成分分析
Numpy scipyで独立成分分析Numpy scipyで独立成分分析
Numpy scipyで独立成分分析
 
Attentionの基礎からTransformerの入門まで
Attentionの基礎からTransformerの入門までAttentionの基礎からTransformerの入門まで
Attentionの基礎からTransformerの入門まで
 
[DL輪読会]GANSynth: Adversarial Neural Audio Synthesis
[DL輪読会]GANSynth: Adversarial Neural Audio Synthesis[DL輪読会]GANSynth: Adversarial Neural Audio Synthesis
[DL輪読会]GANSynth: Adversarial Neural Audio Synthesis
 
異常音検知に対する深層学習適用事例
異常音検知に対する深層学習適用事例異常音検知に対する深層学習適用事例
異常音検知に対する深層学習適用事例
 
Onoma-to-wave: オノマトペを利用した環境音合成手法の提案
Onoma-to-wave: オノマトペを利用した環境音合成手法の提案Onoma-to-wave: オノマトペを利用した環境音合成手法の提案
Onoma-to-wave: オノマトペを利用した環境音合成手法の提案
 
DAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也hare
DAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也hareDAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也hare
DAシンポジウム2019招待講演「深層学習モデルの高速なTraining/InferenceのためのHW/SW技術」 金子紘也hare
 

Similar to 声質変換の概要と最新手法の紹介

CODE FESTIVAL 2015 予選A 解説
CODE FESTIVAL 2015 予選A 解説CODE FESTIVAL 2015 予選A 解説
CODE FESTIVAL 2015 予選A 解説AtCoder Inc.
 
Orb における Cassandra への取り組み
Orb における Cassandra への取り組みOrb における Cassandra への取り組み
Orb における Cassandra への取り組みOrb, Inc.
 
0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD Screen
0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD Screen0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD Screen
0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD ScreenShawn Lee
 
20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤ICT_CONNECT_21
 
20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤ICT_CONNECT_21
 
Attention-Based Adaptive Selection of Operations for Image Restoration in the...
Attention-Based Adaptive Selection of Operations for Image Restoration in the...Attention-Based Adaptive Selection of Operations for Image Restoration in the...
Attention-Based Adaptive Selection of Operations for Image Restoration in the...MasanoriSuganuma
 
Tensorflow and python : fault detection system - PyCon Taiwan 2017
Tensorflow and python : fault detection system - PyCon Taiwan 2017Tensorflow and python : fault detection system - PyCon Taiwan 2017
Tensorflow and python : fault detection system - PyCon Taiwan 2017Eric Ahn
 
音響信号に対する異常音検知技術と応用
音響信号に対する異常音検知技術と応用音響信号に対する異常音検知技術と応用
音響信号に対する異常音検知技術と応用Yuma Koizumi
 
Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...
Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...
Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...Santoshi Family
 
Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...
Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...
Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...Ahmed Gad
 
【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints
【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints
【ECCV 2018】CornerNet: Detecting Objects as Paired Keypointscvpaper. challenge
 
Safe Reinforcement Learning
Safe Reinforcement LearningSafe Reinforcement Learning
Safe Reinforcement LearningDongmin Lee
 
Stargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動する
Stargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動するStargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動する
Stargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動するKohei Tokunaga
 
Systems and methods for visual presentation and selection of ivr menu
Systems and methods for visual presentation and selection of ivr menuSystems and methods for visual presentation and selection of ivr menu
Systems and methods for visual presentation and selection of ivr menuTal Lavian Ph.D.
 
A t-out-of-n Redactable Signature Scheme
A t-out-of-n Redactable Signature SchemeA t-out-of-n Redactable Signature Scheme
A t-out-of-n Redactable Signature SchemeMASAYUKITEZUKA1
 

Similar to 声質変換の概要と最新手法の紹介 (20)

Prelude to halide_public
Prelude to halide_publicPrelude to halide_public
Prelude to halide_public
 
CODE FESTIVAL 2015 予選A 解説
CODE FESTIVAL 2015 予選A 解説CODE FESTIVAL 2015 予選A 解説
CODE FESTIVAL 2015 予選A 解説
 
Orb における Cassandra への取り組み
Orb における Cassandra への取り組みOrb における Cassandra への取り組み
Orb における Cassandra への取り組み
 
0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD Screen
0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD Screen0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD Screen
0.47 inch LCD Micro Dispalay 800x600 Resolution RGB Interface LCD Screen
 
20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤
 
20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤20170322_ICON21技術セミナー1_加藤
20170322_ICON21技術セミナー1_加藤
 
Attention-Based Adaptive Selection of Operations for Image Restoration in the...
Attention-Based Adaptive Selection of Operations for Image Restoration in the...Attention-Based Adaptive Selection of Operations for Image Restoration in the...
Attention-Based Adaptive Selection of Operations for Image Restoration in the...
 
Tensorflow and python : fault detection system - PyCon Taiwan 2017
Tensorflow and python : fault detection system - PyCon Taiwan 2017Tensorflow and python : fault detection system - PyCon Taiwan 2017
Tensorflow and python : fault detection system - PyCon Taiwan 2017
 
音響信号に対する異常音検知技術と応用
音響信号に対する異常音検知技術と応用音響信号に対する異常音検知技術と応用
音響信号に対する異常音検知技術と応用
 
2937
29372937
2937
 
Hong.bas
Hong.basHong.bas
Hong.bas
 
Hong.bas
Hong.basHong.bas
Hong.bas
 
Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...
Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...
Linear Algebra Previous Year Questions of Csir Net Mathematical Science and t...
 
Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...
Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...
Introduction to Artificial Neural Networks (ANNs) - Step-by-Step Training & T...
 
Salesforce Big Object 最前線
Salesforce Big Object 最前線Salesforce Big Object 最前線
Salesforce Big Object 最前線
 
【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints
【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints
【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints
 
Safe Reinforcement Learning
Safe Reinforcement LearningSafe Reinforcement Learning
Safe Reinforcement Learning
 
Stargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動する
Stargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動するStargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動する
Stargz Snapshotter: イメージのpullを省略してcontainerdでコンテナを高速に起動する
 
Systems and methods for visual presentation and selection of ivr menu
Systems and methods for visual presentation and selection of ivr menuSystems and methods for visual presentation and selection of ivr menu
Systems and methods for visual presentation and selection of ivr menu
 
A t-out-of-n Redactable Signature Scheme
A t-out-of-n Redactable Signature SchemeA t-out-of-n Redactable Signature Scheme
A t-out-of-n Redactable Signature Scheme
 

More from Kentaro Tachibana

ICASSP2020音声&音響読み会Mellotron
ICASSP2020音声&音響読み会MellotronICASSP2020音声&音響読み会Mellotron
ICASSP2020音声&音響読み会MellotronKentaro Tachibana
 
Interspeech2019読み会 音声生成
Interspeech2019読み会 音声生成Interspeech2019読み会 音声生成
Interspeech2019読み会 音声生成Kentaro Tachibana
 
ICASSP2019 音声&音響読み会 テーマ発表音声生成
ICASSP2019 音声&音響読み会 テーマ発表音声生成ICASSP2019 音声&音響読み会 テーマ発表音声生成
ICASSP2019 音声&音響読み会 テーマ発表音声生成Kentaro Tachibana
 
Icml2018読み会_overview&GANs
Icml2018読み会_overview&GANsIcml2018読み会_overview&GANs
Icml2018読み会_overview&GANsKentaro Tachibana
 
Icassp2018 発表参加報告 FFTNet, Tactron2紹介
Icassp2018 発表参加報告 FFTNet, Tactron2紹介Icassp2018 発表参加報告 FFTNet, Tactron2紹介
Icassp2018 発表参加報告 FFTNet, Tactron2紹介Kentaro Tachibana
 

More from Kentaro Tachibana (6)

ICASSP2020音声&音響読み会Mellotron
ICASSP2020音声&音響読み会MellotronICASSP2020音声&音響読み会Mellotron
ICASSP2020音声&音響読み会Mellotron
 
Interspeech2019読み会 音声生成
Interspeech2019読み会 音声生成Interspeech2019読み会 音声生成
Interspeech2019読み会 音声生成
 
190910 SHIBUYA synapse
190910 SHIBUYA synapse190910 SHIBUYA synapse
190910 SHIBUYA synapse
 
ICASSP2019 音声&音響読み会 テーマ発表音声生成
ICASSP2019 音声&音響読み会 テーマ発表音声生成ICASSP2019 音声&音響読み会 テーマ発表音声生成
ICASSP2019 音声&音響読み会 テーマ発表音声生成
 
Icml2018読み会_overview&GANs
Icml2018読み会_overview&GANsIcml2018読み会_overview&GANs
Icml2018読み会_overview&GANs
 
Icassp2018 発表参加報告 FFTNet, Tactron2紹介
Icassp2018 発表参加報告 FFTNet, Tactron2紹介Icassp2018 発表参加報告 FFTNet, Tactron2紹介
Icassp2018 発表参加報告 FFTNet, Tactron2紹介
 

Recently uploaded

Formation of low mass protostars and their circumstellar disks
Formation of low mass protostars and their circumstellar disksFormation of low mass protostars and their circumstellar disks
Formation of low mass protostars and their circumstellar disksSérgio Sacani
 
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...Sérgio Sacani
 
CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service 🪡
CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service  🪡CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service  🪡
CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service 🪡anilsa9823
 
Pests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdfPests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdfPirithiRaju
 
Green chemistry and Sustainable development.pptx
Green chemistry  and Sustainable development.pptxGreen chemistry  and Sustainable development.pptx
Green chemistry and Sustainable development.pptxRajatChauhan518211
 
Bentham & Hooker's Classification. along with the merits and demerits of the ...
Bentham & Hooker's Classification. along with the merits and demerits of the ...Bentham & Hooker's Classification. along with the merits and demerits of the ...
Bentham & Hooker's Classification. along with the merits and demerits of the ...Nistarini College, Purulia (W.B) India
 
All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...
All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...
All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...Sérgio Sacani
 
Spermiogenesis or Spermateleosis or metamorphosis of spermatid
Spermiogenesis or Spermateleosis or metamorphosis of spermatidSpermiogenesis or Spermateleosis or metamorphosis of spermatid
Spermiogenesis or Spermateleosis or metamorphosis of spermatidSarthak Sekhar Mondal
 
Botany 4th semester file By Sumit Kumar yadav.pdf
Botany 4th semester file By Sumit Kumar yadav.pdfBotany 4th semester file By Sumit Kumar yadav.pdf
Botany 4th semester file By Sumit Kumar yadav.pdfSumit Kumar yadav
 
Botany krishna series 2nd semester Only Mcq type questions
Botany krishna series 2nd semester Only Mcq type questionsBotany krishna series 2nd semester Only Mcq type questions
Botany krishna series 2nd semester Only Mcq type questionsSumit Kumar yadav
 
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...jana861314
 
Nightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43b
Nightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43bNightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43b
Nightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43bSérgio Sacani
 
TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...
TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...
TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...ssifa0344
 
Grafana in space: Monitoring Japan's SLIM moon lander in real time
Grafana in space: Monitoring Japan's SLIM moon lander  in real timeGrafana in space: Monitoring Japan's SLIM moon lander  in real time
Grafana in space: Monitoring Japan's SLIM moon lander in real timeSatoshi NAKAHIRA
 
SOLUBLE PATTERN RECOGNITION RECEPTORS.pptx
SOLUBLE PATTERN RECOGNITION RECEPTORS.pptxSOLUBLE PATTERN RECOGNITION RECEPTORS.pptx
SOLUBLE PATTERN RECOGNITION RECEPTORS.pptxkessiyaTpeter
 
Botany 4th semester series (krishna).pdf
Botany 4th semester series (krishna).pdfBotany 4th semester series (krishna).pdf
Botany 4th semester series (krishna).pdfSumit Kumar yadav
 
Nanoparticles synthesis and characterization​ ​
Nanoparticles synthesis and characterization​  ​Nanoparticles synthesis and characterization​  ​
Nanoparticles synthesis and characterization​ ​kaibalyasahoo82800
 
Isotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on IoIsotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on IoSérgio Sacani
 
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...anilsa9823
 

Recently uploaded (20)

Formation of low mass protostars and their circumstellar disks
Formation of low mass protostars and their circumstellar disksFormation of low mass protostars and their circumstellar disks
Formation of low mass protostars and their circumstellar disks
 
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
 
CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service 🪡
CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service  🪡CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service  🪡
CALL ON ➥8923113531 🔝Call Girls Kesar Bagh Lucknow best Night Fun service 🪡
 
9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service
9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service
9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service
 
Pests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdfPests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdf
 
Green chemistry and Sustainable development.pptx
Green chemistry  and Sustainable development.pptxGreen chemistry  and Sustainable development.pptx
Green chemistry and Sustainable development.pptx
 
Bentham & Hooker's Classification. along with the merits and demerits of the ...
Bentham & Hooker's Classification. along with the merits and demerits of the ...Bentham & Hooker's Classification. along with the merits and demerits of the ...
Bentham & Hooker's Classification. along with the merits and demerits of the ...
 
All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...
All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...
All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...
 
Spermiogenesis or Spermateleosis or metamorphosis of spermatid
Spermiogenesis or Spermateleosis or metamorphosis of spermatidSpermiogenesis or Spermateleosis or metamorphosis of spermatid
Spermiogenesis or Spermateleosis or metamorphosis of spermatid
 
Botany 4th semester file By Sumit Kumar yadav.pdf
Botany 4th semester file By Sumit Kumar yadav.pdfBotany 4th semester file By Sumit Kumar yadav.pdf
Botany 4th semester file By Sumit Kumar yadav.pdf
 
Botany krishna series 2nd semester Only Mcq type questions
Botany krishna series 2nd semester Only Mcq type questionsBotany krishna series 2nd semester Only Mcq type questions
Botany krishna series 2nd semester Only Mcq type questions
 
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
 
Nightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43b
Nightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43bNightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43b
Nightside clouds and disequilibrium chemistry on the hot Jupiter WASP-43b
 
TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...
TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...
TEST BANK For Radiologic Science for Technologists, 12th Edition by Stewart C...
 
Grafana in space: Monitoring Japan's SLIM moon lander in real time
Grafana in space: Monitoring Japan's SLIM moon lander  in real timeGrafana in space: Monitoring Japan's SLIM moon lander  in real time
Grafana in space: Monitoring Japan's SLIM moon lander in real time
 
SOLUBLE PATTERN RECOGNITION RECEPTORS.pptx
SOLUBLE PATTERN RECOGNITION RECEPTORS.pptxSOLUBLE PATTERN RECOGNITION RECEPTORS.pptx
SOLUBLE PATTERN RECOGNITION RECEPTORS.pptx
 
Botany 4th semester series (krishna).pdf
Botany 4th semester series (krishna).pdfBotany 4th semester series (krishna).pdf
Botany 4th semester series (krishna).pdf
 
Nanoparticles synthesis and characterization​ ​
Nanoparticles synthesis and characterization​  ​Nanoparticles synthesis and characterization​  ​
Nanoparticles synthesis and characterization​ ​
 
Isotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on IoIsotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on Io
 
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
 

声質変換の概要と最新手法の紹介

  • 1. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential
  • 2. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential n G n n n - n A C 2
  • 3. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential n a / []eb A D a C / D 4 1 /0 , 1 6 3
  • 4. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential n I N ) n A B B ( ( ( ( (
  • 5. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential 5 3 0 2 . /3 0 7 3 1 . 0 0 7 3
  • 6. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential 6 3 0 2 . /3 0 7 3 1 . 0 0 7 3 :
  • 7. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential n A B A 7 B ( A ) ( ; F0) ( ; bap) B
  • 8. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential 8 (F0 ) bap • → Vocoder • • STRAIGHT [Kawahara+; ’99] • WORLD [Morise+; ’16]
  • 9. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential Vocoder 9
  • 10. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential vocoder 10 F0bap F0bap F0bap 1 frame frame Frame
  • 11. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential [Abe+; ’90][stylianou+; ’98] n A B 11 F0bap F0bap GMM DNN
  • 12. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential [Abe+; ’90][stylianou+; ’98] n B 12 AF0bap AF0bap GMM DNN • F0, bap → A • F0 bap
  • 13. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential 13 Parallel-data B A frame A B
  • 14. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential o s e P P • 6C C E C AA 6-- ( y • 6-- K K E 2; ] p e d r aKvt 6-- ( 3 E AA A ; G • g K V P[P PN g kO • E AA A ; G P Po - A 1 70 C + )8 i h ced • nQc i h ʻ] d l ]SP • nQc 6 6 . 7 ; 2CE; + )8 14
  • 15. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential ig l hV vNVQ] V • 3 7 7 E C 7;6 Nk 16 6 6 6Nyo • Ns V PV cK • A6 6 6 6 V i + 7 - .6 G n a N [] • n a r pdNʻ O • t V e [n 32 3 , E6 0 G 15
  • 16. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential Voice Conversion Challenge 2016 n n 7 7 7 n 5 5 n 01 16
  • 17. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential Results of listening tests in VCC 2016 17 cf. http://vc-challenge.org/vcc2016/summary.html
  • 18. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential ig l hV vNVQ] V • 3 7 7 E C 7;6 Nk 16 6 6 6Nyo • Ns V PV cK • A6 6 6 6 V i + 7 - .6 G n a N [] • n a r pdNʻ O • t V e [n 32 3 , E6 0 G 18
  • 19. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential CycleGAN [Zhu+; ’17] n 19 cf. https://junyanz.github.io/CycleGAN/
  • 20. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential CycleGAN [Zhu+; ’17] n n Forward-inverse mapping Inverse-forward mapping GX→Y GY→X G L real/fake loss [Kaneko+; ‘17] M mapping loss
  • 21. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential CycleGAN [Zhu+; ’17] n n 21 Forward-inverse mapping Inverse-forward mapping GX→Y adversarial loss
  • 22. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential CycleGAN [Zhu+; ’17] n n 22 Forward-inverse mapping Inverse-forward mapping = "#~%&'(' # log,- . + "0~%&'(' 0 log 1 − ,- 34→- . GY→X adversarial loss
  • 23. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential CycleGAN [Zhu+; ’17] n n 23 Forward-inverse mapping Inverse-forward mapping L1loss
  • 24. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential CycleGAN [Zhu+; ’17] n n 24 Forward-inverse mapping Inverse-forward mapping λcyc 10.0
  • 25. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential CycleGAN parallel-data-free [Kaneko+, ’17] n NG NG n C 25 CycleGAN copy A A A A A t1 t2 tTbap bap bap bap bap F0 F0 F0 F0 F0 bap bap bap bap bap F0 F0 F0 F0 F0 A A A A A t1 t2 tT
  • 26. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential VC . 1 1 2 r • d c U R l U • c U • t pt t G l em - 1 1 a ) ( A . (1 1 ( l • y sv cG X • g I ni L UI o 26
  • 27. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential Network architecture n . / - . / n : . / / . / / / . 27
  • 28. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential ig l hV vNVQ] V • 3 7 7 E C 7;6 Nk 16 6 6 6Nyo • Ns V PV cK • A6 6 6 6 V i + 7 - .6 G n a N [] • n a r pdNʻ O • t V e [n 32 3 , E6 0 G 28
  • 29. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential Variational autoencoder (VAE) [Hinton+; '06] n z 29 x Encoder qθ(z|X) Decoder pθ(X|z) z !" # $; 0, 1 Input feature Generated feature
  • 30. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential VQ-VAE [van den Oord+, ‘17] n - -( ) E V A n 30 x Encoder p(ze(x)|x) Decoder p(x|zq(x)) ze(x) !" A A e1 e2 e3 eK zq(x) x LQ loss VQ loss Encoder loss
  • 31. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential VQ-VAE n [van den Oord+; ’16] t v o G a r • N x h G d • λ W l lg d • l m r e 31 ! " # = % &'( ) * +&|+&-),+&-)/0, ⋯ +&-0, # λ : d lg c d , " = +(, +0, ⋯ +&-0
  • 32. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential VQ-VAE [van den Oord+, ’17] n 32 Encoder WaveNet ze(x) e1 e2 e3 eK zq(x) id • zq(x) id • ze(x) zq(x) id • zq(x) ( ) https://avdnoord.github.io/homepage/vqvae/
  • 33. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential VQ-VAE [van den Oord+; ‘17] n 33 Encoder WaveNet ze(x) e1 e2 e3 eK zq(x) id cf. https://www.slideshare.net/YukiSaito8/saito18sp03 • zq(x) id • ze(x) zq(x) id • zq(x) ( ) https://avdnoord.github.io/homepage/vqvae/
  • 34. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential n A n C - 34
  • 35. Copyright © DeNA Co.,Ltd. All Rights Reserved. Strictly confidential H9J J JZJ d.-I 6 9J J JZJ 7 J[]MJ 9J[][N JVM 0 MN 1 N NRPVN e N[ Z]L ]ZRVP [XNNL ZNXZN[NV J RWV[ ][RVP J XR L JMJX R N RUN OZNY]NVLa [UWW RVP JVM JV RV[ JV JVNW][ OZNY]NVLa KJ[NM 4 N ZJL RWV W[[RKTN ZWTN WO J ZNX R R N [ Z]L ]ZN RV [W]VM[ f XNNL 1WUU]VRLJ RWV , XX -, , ... H WZR[N c +I WZR[N 4 FWSWUWZR JVM 9 bJ J eD :2 J WLWMNZ KJ[NM RP Y]JTR a [XNNL [aV N[R[ [a[ NU OWZ ZNJT RUN JXXTRLJ RWV[ f 73713 ZJV[JL RWV[ WV RVOWZUJ RWV JVM [a[ NU[ WT 3.. 2 VW , XX -,, --) + H0KN . I 0KN JSJU]ZJ 9 RSJVW JVM 6 9] JKJZJ eCWRLN LWV NZ[RWV ZW]P NL WZ Y]JV RbJ RWV f , ,+ H[ aTRJVW] c.-I F aTRJVW] 1JXXh JVM 3 W]TRVN[ e1WV RV]W][ XZWKJKRTR[ RL ZJV[OWZU OWZ WRLN LWV NZ[RWV f ( ) ..- H9JVNSW ,I A 9JVNSW JVM 6 9JUNWSJ f JZJTTNT 2J J 4ZNN CWRLN 1WV NZ[RWV [RVP 1aLTN 1WV[R[ NV 0M NZ[JZRJT N WZS[ f JZER , HG ] c ,I 8 F G ] A JZS 7[WTJ JVM 0 0 3OZW[ e VXJRZNM RUJPN W RUJPN ZJV[TJ RWV ][RVP LaLTN LWV[R[ NV JM NZ[JZRJT VN WZS[ f H6RV WV c +I 5 3 6RV WV JVM JTJS ] MRVW e NM]LRVP N MRUNV[RWVJTR a WO MJ J R VN]ZJT VN WZS[ f ,-+ ) , + H JV MNV WZM c ,I 0 JV MNV WZM JVM CRVaJT[ e N]ZJT MR[LZN N ZNXZN[NV J RWV TNJZVRVP f 7V XX +( . +( - , H JV MNV WZM d +I 0 JV 2NV WZM 2RNTNUJV 6 GNV 9 RUWVaJV CRVaJT[ 0 5ZJ N[ JVM 9 9J ]SL]WPT] eDJ NVN 0 PNVNZJ R N UWMNT OWZ ZJ J]MRW f 35