"In March 2012, Japan’s leading mobile operator, NTT DOCOMO, introduced Shabette Concier, an advanced voice-activated personal agent service that enables customers to intuitively and directly operate services and smartphone features with voice commands. Millions of DOCOMO's subscribers are now using this service.
This session explains Shabette Concier's distributed speech recognition architecture, and dialogue-understanding system design, with machine learning technologies and large-scale database systems. Learn why DOCOMO chose the AWS cloud and how DOCOMO engineers overcame all the difficulties from CEO-imposed time constraints, unexpectedly rapid service growth, usage spikes driven by marketing campaigns, and internal resistance to the use of cloud services. The session concludes with lessons learned from a telco's large-scale service development of a mobile app with the AWS cloud."
11. Who is DOCOMO?
•
$61 million subscribers in Japan
•
$45 billion revenue (2013)
3
Friday, November 15, 13
12. Who is DOCOMO?
•
$61 million subscribers in Japan
•
$45 billion revenue (2013)
Next to China Mobile, Verizon, and
AT&T
3
Friday, November 15, 13
13. Who is DOCOMO?
•
$61 million subscribers in Japan
•
$45 billion revenue (2013)
Next to China Mobile, Verizon, and
AT&T
•
“i-mode” innovator (1999)
3
Friday, November 15, 13
14. Revenue Structure (FY2012)
Feature Phone Era
Data Stage
Voice Stage
Service
Cloud
Packet
Network
Voice
Network
Device
4
Friday, November 15, 13
15. Revenue Structure (FY2012)
Feature Phone Era
Data Stage
Voice Stage
Smart Phone
Service
Cloud
Service $5.5 Billion
Packet
Network
Data: $19 Billion
Voice
Network
Voice: $13 Billion
Device
4
Friday, November 15, 13
29. “Shabette-Concier” Voice agent service
Concier
=
=
How may I help you?
Shabette
Voice
Concierge
•
•
•
•
Launched Mar. 1, 2012
Over 40 services in it
Including chatting
10 million users
12
Friday, November 15, 13
55. In Wannabe Skunkworks (2010)
Maybe, it’s
high time to make
a voice agent.
But, as a stealth product.
23
Friday, November 15, 13
56. In Wannabe Skunkworks (2010)
Maybe, it’s
high time to make
a voice agent.
But, as a stealth product.
Yep, we have lot of expertise & data for
speech recognition.
23
Friday, November 15, 13
57. In Wannabe Skunkworks (2010)
Just out of
curiosity,
I’d like to use Public
Cloud.
Maybe, it’s
high time to make
a voice agent.
But, as a stealth product.
Yep, we have lot of expertise & data for
speech recognition.
23
Friday, November 15, 13
58. In Wannabe Skunkworks (2010)
Just out of
curiosity,
I’d like to use Public
Cloud.
Maybe, it’s
high time to make
a voice agent.
But, as a stealth product.
Yep, we have lot of expertise & data for
speech recognition.
23
Friday, November 15, 13
59. Distributed Speech Recognition (2006)
First Implementation of ETSI AURORA Project (2000-2003)
16 Khz Sampling MFCC feature vectors --> 5.6 Kbps
24
Friday, November 15, 13
60. Distributed Speech Recognition (2006)
Mobile Terminal
Dictionary
Server
Feature
Extraction
Encode
Decode
Rec.
Engine
Result
First Implementation of ETSI AURORA Project (2000-2003)
16 Khz Sampling MFCC feature vectors --> 5.6 Kbps
24
Friday, November 15, 13
64. The paradigm shift from ‘Search’
Microsoft Bing
Google Voice Search
26
Friday, November 15, 13
65. A Small Team in Wannabe Skunkworks (2010-2012)
Product owner
Development
promoter
Core engineer
Agile
Development
Core engineer
Development
promoter
27
Friday, November 15, 13
66. A Small Team in Wannabe Skunkworks (2010-2012)
Product owner
Development
promoter
Core engineer
Agile
Development
Core engineer
Development
promoter
27
Friday, November 15, 13
69. Basic Architecture 2010
(inspired by Microsoft Bing, and Google Voice Search)
text
Text to speech
contents
Voice
text
text
contents
Voice
Recognition
Logging
Friday, November 15, 13
Task
Recognition
Logging
Service
Providers’ DB
28
70. History of Voice Agent
5/2011
Stealth Product
Data Center
29
Friday, November 15, 13
(Hybrid with other Cloud)
71. History of Voice Agent
5/2011
Stealth Product
3/2012
Version 1
Mar.∼
Data Center
Friday, November 15, 13
Public
Cloud
29
(Hybrid with other Cloud)
72. History of Voice Agent
5/2011
Stealth Product
3/2012
Version 1
Mar.∼
Data Center
Friday, November 15, 13
Public
Cloud
June
AWS
NC-reg
29
(Hybrid with other Cloud)
73. History of Voice Agent
5/2011
Stealth Product
3/2012
Version 1
Mar.∼
Data Center
Friday, November 15, 13
Public
Cloud
11/2012
Version 2
June
AWS
NC-reg
Sept.
AWS
Tokyo-reg.
(Hybrid with other Cloud)
29
74. Voice Recognition
I want...
Speech Recognition
Text
Speech
Acoustic
Model
Machine Learning
Speech
Data
Lexicon
Language
Model
Machine Learning
Language
Data
30
Friday, November 15, 13
75. Voice Recognition
I want...
Speech Recognition
Text
Speech
Acoustic
Model
Machine Learning
Speech
Data
Lexicon
Language
Model
Machine Learning
Language
Data
BELIEF: More data usually beats better algorithms.
30
Friday, November 15, 13
76. Task Recognition and Service Flow
Restaurant
Lexicon
Task Corpus
Dictionary
I want...
Tokenizer
Text
Abstractor
Contents
Feature
extractor
Abstraction
Dictionary
MC-SVM
Classifier
Query ext.
Find a good Italian restaurant in Palo Alto
31
Friday, November 15, 13
77. Task Recognition and Service Flow
Restaurant
Lexicon
Task Corpus
Dictionary
I want...
Tokenizer
Text
Abstractor
Contents
Feature
extractor
Abstraction
Dictionary
MC-SVM
Classifier
Service
Launcher
Search Engine A
Query ext.
Search Engine B
Search Engine C
Find a good Italian restaurant in Palo Alto
PRINCIPLE: Machine learning enhances the service quality.
A better quality service acquires more data.
Friday, November 15, 13
31
78. Version 1 Implementation (Mar. 2012)
Started with two local cloud
providers in Japan, and soon faced
two difficulties:
32
Friday, November 15, 13
79. Version 1 Implementation (Mar. 2012)
Started with two local cloud
providers in Japan, and soon faced
two difficulties:
LB
A
B
32
Friday, November 15, 13
80. Version 1 Implementation (Mar. 2012)
Started with two local cloud
providers in Japan, and soon faced
two difficulties:
LB
A
B
• Scalability in
server count
32
Friday, November 15, 13
81. Version 1 Implementation (Mar. 2012)
Started with two local cloud
providers in Japan, and soon faced
two difficulties:
LB
A
B
• Scalability in
server count
• Inflexible monthly
payment
32
Friday, November 15, 13
84. Moved to AWS (June, 2012)
34
Friday, November 15, 13
85. Moved to AWS (June, 2012)
Scale!
34
Friday, November 15, 13
86. Moved to AWS (June, 2012)
Technology
Platform
Leader
AWS Partners
Innovative
Ecosystem
Scale!
Global User
Footprint
34
Friday, November 15, 13
87. System Architecture (June 2012)
SmartPhone
VPC
ELB
Management
Server
Log Server for VR
(across multiple zones)
Same as
AZ #1
TR Servers
VR Servers
Voice Recognizer(VR)
Availability Zone #1
(across multiple
zones)
LB
ELB
ELB
(across multiple
zones)
Task Recognizer(TR)
Availability Zone #2
Tokenizer
Access Log Servers
Log management system
Availability Zone #3
35
Friday, November 15, 13
88. Killer Design Pattern: Multi-Data Center
Voice Recognition Part
Voice
VPC
Route 53
LB
VR servers
Availability Zone #1
HAProxy
LB
m2.4xlarge
x ~300 !!
VR servers
Availability Zone #2
36
Friday, November 15, 13
89. DOCOMO Skunkworks: Cloud Natives
• For another service,
adopted 21 out of 48
design patterns.
• Among them, typically
used ‘Queuing
Chain’
and
‘Scale
Out’
Pa4erns.
37
Friday, November 15, 13
100. Prepare for Launching Servers
Refresh AMI every day
–VR model has been
changed everyday
Launch AMI for bursty
traffic
–10 min - 30 min to launch
•Loading VR model is
taking time....
42
Friday, November 15, 13
103. Moved twice
• Apr 2012
–Start to plan for moving-out to AWS
• Jun 2012
–Moved to AWS Northern California
• Tokyo did not have enough
instances at that time
• Deploy mixture instance types
(c1.xlarge and m2.4xlarge)
• Sep 2012
–Moved to AWS Tokyo
• Unified to m2.4xlarge
–Improved latency
Friday, November 15, 13
44
105. October, 2012
Kid, your game
is over.
“Status Quo”
DOCOMO
“99.9999%”
Shrine
Skunkworks
Temple
Office of
Inspirations
(not ‘Wannabe’ anymore)
45
Friday, November 15, 13
106. October, 2012
Use our
stable system.
Kid, your game
is over.
“Status Quo”
DOCOMO
“99.9999%”
Shrine
Skunkworks
Temple
Office of
Inspirations
(not ‘Wannabe’ anymore)
45
Friday, November 15, 13
107. October, 2012
Use our
stable system.
Kid, your game
is over.
Move on
quickly!
“Status Quo”
DOCOMO
“99.9999%”
Shrine
Skunkworks
Temple
Office of
Inspirations
(not ‘Wannabe’ anymore)
45
Friday, November 15, 13
108. October, 2012
Catch me
if you can
SCALE!
Use our
stable system.
Kid, your game
is over.
Move on
quickly!
“Status Quo”
DOCOMO
“99.9999%”
Shrine
Skunkworks
Temple
Office of
Inspirations
(not ‘Wannabe’ anymore)
45
Friday, November 15, 13
110. In DOCOMO Skunkworks (2013)
They are SLOW and
pricy. Keep Restless
improvement and our leanstartup culture.
46
Friday, November 15, 13
111. In DOCOMO Skunkworks (2013)
They are SLOW and
pricy. Keep Restless
improvement and our leanstartup culture.
Yep, let’s
continue our system
improvement on the fly.
46
Friday, November 15, 13
112. In DOCOMO Skunkworks (2013)
Hey, leader,
how about multi-lingual
translation services?
They are SLOW and
pricy. Keep Restless
improvement and our leanstartup culture.
Yep, let’s
continue our system
improvement on the fly.
46
Friday, November 15, 13
113. In DOCOMO Skunkworks (2013)
Hey, leader,
how about multi-lingual
translation services?
They are SLOW and
pricy. Keep Restless
improvement and our leanstartup culture.
Yep, let’s
continue our system
improvement on the fly.
46
Friday, November 15, 13
114. Now, we have …
17-Nov-16 27-Nov-16 7-Dec-16 17-Dec-16 27-Dec-16 6-Jan-17 16-Jan-17 26-Jan-17 5-Feb-17 15-Feb-17 25-Feb-17 7-Mar-17 17-Mar-17 27-Mar-17 6-Apr-17 16-Apr-17 26-Apr-17 6-May-17 16-May-17 26-May-17 5-Jun-17 15-Jun-17 25-Jun-17
17-Nov-12
5-Jul-17
15-Jul-17 25-Jul-17
4-Aug-17 14-Aug-17 24-Aug-17
Number of Speech
• 10 million terminals pre-installed
• More than 4M distinct user accesses
• 1M accesses / day
47
Friday, November 15, 13
115. Other Tips
• Instance Management
✓Simple DB as a key-value
store
• Log management
✓Enormous logs from servers
✓Direct write to S3
• Multi-AZ
✓HA and act-act configuration
48
Friday, November 15, 13
116. Thanks to the people who gave great
lessons to ‘DOCOMO Skunkworks.’
49
Friday, November 15, 13
117. Thanks to the people who gave great
lessons to ‘DOCOMO Skunkworks.’
in changing our development style to
49
Friday, November 15, 13
118. Thanks to the people who gave great
lessons to ‘DOCOMO Skunkworks.’
in changing our development style to
“Deploy first, think later,”
49
Friday, November 15, 13
119. Thanks to the people who gave great
lessons to ‘DOCOMO Skunkworks.’
in changing our development style to
“Deploy first, think later,”
49
Friday, November 15, 13
120. Thanks to the people who gave great
lessons to ‘DOCOMO Skunkworks.’
in changing our development style to
“Deploy first, think later,”
in bringing us a high performance culture,
and
49
Friday, November 15, 13
121. Thanks to the people who gave great
lessons to ‘DOCOMO Skunkworks.’
in changing our development style to
“Deploy first, think later,”
in bringing us a high performance culture,
and
49
Friday, November 15, 13
122. Thanks to the people who gave great
lessons to ‘DOCOMO Skunkworks.’
in changing our development style to
“Deploy first, think later,”
in bringing us a high performance culture,
and
in becoming “Cloud Natives.”
49
Friday, November 15, 13
123. Please give us your feedback on this
presentation
MBL202
As a thank you, we will select prize
winners daily for completed surveys!
Thank You
50
Friday, November 15, 13