Transcript Tutorial on Sounds of Silence" - B. Yegnanarayana

Sounds of Silence
The Challenge for AI
B.Yegnanarayana
Speech and Vision Lab
Dept. of CS&E, IIT Madras
Goal of Artificial Intelligence
The urge is to make machines more intelligent
But in the process we are doing the opposite
Why?
Because we are only storing and manipulating data
What is INTELLIGENCE?
It is not simply manipulation of data
Intelligence of human beings
Capture, associate and retrieve patterns
Examples
Signatures, face recognition, video
< Prev
Next >
Why AI is difficult: Some examples
Representation of 1D, 2D, 3D data
Contrast
Reading vs writing
Listening vs speaking
Looking vs sketching
Watching vs doing
Recognition vs synthesis
Key is learning
Development of motor control is slow
Intelligent activity
Involves linking key features/concepts/ideas
< Prev
Next >
Intelligence vs Information
Creating an environment for intelligent activity
Current methods do exactly the opposite
We present more data more frequently
No scope for acquiring implicit pattern behavior
Confusion between knowledge and information
Knowledge society or ignorance society
Filling up mind with data is like filling up silence
Intelligence is in capturing the sounds of silence
< Prev
Next >
Sounds of Silence
Significance of silence
In cartoons and string of characters
Examples of silence in speech sounds
A sufficient cue
A necessary cue
Illustrations from continuous speech
Waveform, residual and impulses
Different speakers
Different languages
Perception of Sounds of Silence
Human ability and machine's inability - Why?
< Prev
Next >
Architectural Mismatch
Machines take mostly silence and may ignore signal
Representation
Machine
Human
Pixels/samples
(mostly silence data)
Symbols and
interrelations
(ignores silence)
Multiple
(neurons)
Processor
Single
Processing
Sequential
(local)
< Prev
Parallel and
distributed
(local and global)
Next >
Nature of AI Problems
Data
String of
Characters
Information
Words/
Sound Units
Speech
Sequence of
Samples
Formants &
Pitch
Image
Array of
Pixels
Objects &
Interrelation
Natural
Language
Knowledge Intelligence
Rules
Message
(Syntax)
Intonation &
Duration
Message
(language
constraints)
Rules to
Message
form picture
Decision
making
< Prev
Next >
Illustrations from Speech
• Nature of speech production and perception
• Challenges in speech recognition, synthesis, and
speaker recognition
• Why they are difficult for machine and easy for
us?
• Due to our ability to capture sounds of silence
< Prev
Next >
Characteristics of Human Problem
Solving
• Computing sounds of silence
• Essentially pattern processing instead of data
processing
• Integration of local and global patterns
• Delayed decisions
• Nonuniqueness of solutions
< Prev
Next >
Architectural Features of Possible New
Models
Need to move from
• Deterministic computation to decision logic
• Sequential processing to PDP
• Set of equations to set of inequalities
• Problem solving to learning
• Data processing to multidimensional pattern
processing
< Prev
Next >
Conclusions
• Powerful computers need not solve intelligent
problems
• Finer sampling need not result in good solutions
• Two interesting problems: Video processing and
dictation machine
• The challenge is computing the sounds of silence
• Unless we watch, the technology may destroy itself
by exposing its limitations.
• Dont forget that it is always the human being is the
reference not the machine for intellectual abilities.
< Prev
Next >
Thank you
< Prev
Next >
Back
< Prev
Next >
Some Illustrations of “Sounds of Silence”
Silence:
A sufficient cue
for stop
consonant
perception
slit
split
s_lit
30 ms
150 ms
600 ms
Silence:
A necessary cue
for stop
consonant
perception
sha
shka
10 ms
100 ms
Back
< Prev
Next >
Some Illustrations of “Sounds of Silence”
Signal
Residual
Instants
More examples : Signal, residual and instants
Some more examples : Signal, residual and instants
Back
< Prev
Next >
Speech Production Mechanism
Back
< Prev
Next >
Still Frame
Video Sequence
Less
noise
More
noise
< Prev
Next >
Still Frame
Video Sequence
Less
noise
More
noise
< Prev
Next >
Back
< Prev
< Prev