Upload
bharath-vaishnov
View
123
Download
2
Embed Size (px)
DESCRIPTION
concepts of speech recognition made easy
Citation preview
BIOMETRICS
• USES PHYSICAL OR BEHAVIOURAL CHARACTERISTICS
• DIFFERENT TYPES OF BIOMETRICS
• BIOMETRIC SYSTEM
VOICE RECOGNITION
• RECOGNIZES PEOPLE FROM THEIR VOICES
• VOICE IS A UNIQUE CHARACTER TRAIT
• VOICE DEPENDS ON MANY FACTORS
• TYPES OF VOICE RECOGNITION
SPEECH AND SPEAKER RECOGNTION
• SPEECH->WHAT?
• SPEAKER->WHO?
SPEAKER RECOGNITION
Speaker Recognition
Classification –Styles of input
Speaker Identification
Speaker Verification
A technology that verifies a speaker’s identity based on the speaker’s voice.
Technology of determining an unknown speaker's identity.
Classification – Styles of input
Speaker Recognition
Text Independent Text Prompted
Text Dependent
Fixed Phrase Fixed Phrase
Speaker #1 Speaker #n
Speaker #2
...
Feature Extractor
Hypotheses space
Speaker Modeling
Under preparation
Hypotheses space
Hypothesis representation
Desired hypothesis
define
SearchTraining
examples
Best fit?
Markov Models
• Model to capture extracted voice features
• Probablistic process
• state diagrams- states , transitions
• applied in weather forecasts,dna modelling x — hidden states
y — observable outputsa — transition probabilitiesb — output probabilities
• Coin toss method
• 2 coins A,B: head or tail of any one may appear which is a probability
• feature vectors correspond to head /tail
• state corresponds to the coins A/B
Hidden markov models• Observations are probablistic functions
of states: urn and ball model
• main elements of hmm
• {1,2…N}-individual states ,the initial state time being qt
• {v1,v1…vm}-observation symbols
• state -transition probability
• Observation symbol probability distribution
• B = {bik = P(ok | qi)}
• initial state distribution
• Π = {pi = P(qi at t=1)}.
• System is given by:
• F= (A, B, Π).
Baye’s rule
P(A).P(B|A) = P(B).P(A|B)
P(Fj|OT) = P(OT|Fj).P(Fj) P(OT)
Fj speaker model
OT feature vector of test utterance
SPEAKER IDENTIFICATION
The model with maximum probability of P(Fj|OT) is identified as speaker.
Speaker X
Speaker Y
Speaker Z
Speaker with max probability
Identified
Feature vectorsO1,O2,….OT
SPEAKER VERIFICATION
Target model FA
Background models FB
P(FA|OT) P(OT|FA).P(FA)/P(OT)
P(FB|OT) P(OT|FB).P(FB)/P(OT)
taking log
X = log[P(OT|FA)] – log[P(OT|FB)]
Speaker Y
Imposter 1
Imposter 2
Imposter 3
Feature vectorsO1,O2,..
OT
X
X>=0, acceptX<0, reject
Speaker verification
APPLICATIONS OF SPEAKER RECOGNITION
• USED TO SECURE OUR COMPUTER PASSWORDS
• USED IN ATM’S
• POTENTIAL APLLICATIONS
ADVANTAGES OF SPEAKER RECOGNITION
• IN THE WORLD OF COMPUTER AND INTERNET
• IN THE FIELD OF CREDIT CARD,DEBIT CARDS AND ATM
• POTENTIAL COSTUMERS
DISADVANTAGES OF SPEAKER RECOGNITION
• LANGUAGE PROBLEMS
• CAN BE EASILY OPENED BY MIMICRYING
• IN CASE OF TEMPERORY USE WE CANT HAVE THIS SYSTEM
• TECHNICAL AND HARDWARE PROBLEMS