SLIDE 1
What’s a “wake word”?
- Wake word
- Hot word
- Offline
- Code runs on
CPU/DSP/MCU
- 7x24
- Always listening
- One shot
understanding
- Online
- Code runs on cloud
- On Demand
- Explicit permission
Alexa OK Google Hey Siri
Developing Your Own Wake Word Engine Just Like Alexa and OK Google - - PowerPoint PPT Presentation
Developing Your Own Wake Word Engine Just Like Alexa and OK Google - - PowerPoint PPT Presentation
Developing Your Own Wake Word Engine Just Like Alexa and OK Google Xuchen Yao, CEO, KITT.AI Guoguo Chen, CTO, KITT.AI Whats a wake word? Alexa whats the weather today? OK Google Hey Siri Wake word One shot
SLIDE 2
SLIDE 3
Conversational UI Pipeline
wake up device speech text text understanding dialogue management text speech text voice
SLIDE 4
a customizable hotword detection engine a.k.a: deep neural network in 2MB of RAM hotword.io video blog
SLIDE 5
SLIDE 6
10,000+ developers, 7000+ unique hotwords
Who’s using it (released 5/2016)
Dominating developer community for hotword detection
SLIDE 7
Use Cases
SLIDE 8
#1 Hotword: Smart Mirror
https://github.com/evancohen/smart-mirror (credits to Evan Cohen) video link
SLIDE 9
Command & Control: GoPiGo
(credits to Paul Matz) video link
SLIDE 10
Project RePL
(credits to Chris Burns) video link
SLIDE 11
Conversational UI Pipeline
wake up device speech text text understanding dialogue management text speech text voice Speech Pipeline
SLIDE 12
Voice Microphone Array Wake Word Detection Speech Recognition local
- Close talking
- Far field (3-9
feet)
- 2, 4, or 6
microphones
- Linear/circular
cloud/local
- Voice Activity
Detection
- Auto Gain
Control
- Fast response
(0.1 second)
- High accuracy
- Adaptive Echo
Cancellation
- Beam forming
- IBM/Microsoft/Nua
nce/Google
- Alexa Voice Service
- Kaldi
- PocketSphinx
- HTK
- Command & Control
- Language
Understanding
- Telephone
(8KHz Sampling)
- Others (16KHz)
- Noises: TV,
radio, street, café, car, music
- Pitch: children,
adults, senior
- Accent:
US/UK/Europe/ Asian…
Speech Pipeline
SLIDE 13
Supported Platforms and Wrappers
- Raspberry Pi
- Mac OS X
- iPhone/iPad/iPod
- x86/64bit Ubuntu
- Android
- Pine 64
- Intel Edison
- Samsung Artik
- Allwinner R-series
- Ingenic X1000
- Rockchip
SLIDE 14
Personal vs. Universal models
Personal Universal Voice samples needed 3 At least 1500 Speaker-independent No Yes Speaker-specific Sort of No Robust against noise No Yes Free Yes No Time needed Immediately 2 weeks
SLIDE 15
Customizing a universal model
define hotword collect voice train a model deliver & evaluate deploy to beta users ship & success collect voice from device hotword web API Iterate & Improve desired performance: >90% detection rate <= 3 false alarms in 24 hours
SLIDE 16
Science behind wake word
SLIDE 17
Challenges
- High detection rate
- Low false alarm
- Efficient: detect every 0.1
second
- Small RAM: <2MB
- Too much ambiguity, not
much context
Is this “Alexa”?
short window longer window
SLIDE 18
Existing Algorithm
SLIDE 19
Existing Algorithm
SLIDE 20
Existing Algorithm
- Advantage:
–Simplified pipeline –Simplified decoder
- Disadvantage:
–Massive hotword specific training data
SLIDE 21
Possible Ways to Improve
- Data augmentation
– Adding noise – Adding reverberation – And so on…
- riginal
add noise add noise and reverberation
SLIDE 22
Possible Ways to Improve
- Network models
– Model selection
- Feedforward models? Recurrent models?
– Model compression
- 32-bit float 16-bit float 8-bit integer
- Parameters with small absolute value
SLIDE 23
Possible Ways to Improve
- Decoder redesigning
– Modeling smaller units
- Syllables, phones, etc
– False alarm suppression
- Additional classifier?
SLIDE 24
Training with Tesla K20/K80
- Positive data
– 1,500 hotword samples
- Negative data
– Thousands of hours of speech
- Training time
– Half a day with 4 K80 GPUs
SLIDE 25
Software Architecture
Frontend Backend
SLIDE 26
KITT.AI Scientific Computing
Deep Learning Cloud Devices Production Cloud
Traffic ELB
Content
Websocket audio, msg
HTTPs
Message Queue
Data Training Model Deploy
SLIDE 27