Voice Biometrics

1
Voice Biometrics
2
Topics covered

• History
• Speech vs Voice
• How it works
• Algorithms
• Problems with VR
• Modern Applications
• Future Why VR is good

3
History of Speech Recognition

• 1870
• Alexander Bell invented the telephone
• Wanted to help people with hearing difficulties
• 1922
• REX a toy dog that would come out of its house
  when a certain amount of acoustic energy was
  received

4
History (contd)

• 1936
• Bell Labs invented the voice synthesizer which is
  better known today as a computer voice
• 1960s
• Bell Labs invented a speech recognition system
  with a 10-word vocabulary

5
History (contd)

• 1970s
• Threshold Technology released the VIP 100 System
• First system with a 1000-word vocabulary
• 1978
• Texas Instruments released the Speak and Spell, a
  toy that spoke words you asked it to

6
History (contd)

• 1989
• 486 Processor voice recognition systems became
  more powerful because of processors were more
  powerful
• Less expensive because personal computers were
  becoming more popular

7
History (contd)

• 1997
• Dragon Systems released Naturally Speaking,
  which was the first continuous speech recognition
  system with a large vocabulary
• IBM soon released VoiceType and ViaVoice, which
  are comparable systems to Naturally Speaking

8
History (contd)

• 2004
• Dragon Systems Naturally Speaking 8
  Continuous and Natural Dictation at 160 words per
  minute
• 250,000 word vocabulary, with capabilities to add
  more

9
Speech vs Voice Recognition

• Speech Recognition
• Recognizing the voice of a wide range of users
• Used to recognize words, not identification
• Uses include dictation software for programs such
  as Word, audio commands in modern automobiles,
  telephone menus, etc
• Voice Recognition
• Recognizing the voice of selected users
• Used to identify and/or authenticate users
• Uses include personal, corporate and government
  security.

10
Speech Recognition

• 3 Types
• Discreet word
• Connected word
• Continuous speech

11
Discreet Word

• Requires user to pause between words so software
  can recognize word.
• Works well for systems that only require one word
  responses, i.e. telephone menu with yes/no
  answers.
• Useless otherwise

12
Connected Word

• Similar to discreet word, but allow for multiple
  word phrases or commands.
• User must speak every word clearly and concisely
  so software can recognize each word.
• Improvement over discreet word, but still
  inefficient.

13
Continuous Speech

• Most natural and user friendly
• Very difficult to implement
• Can speak as though having a normal conversation
  with another person.

14
How speech recognition works

• User speaks into microphone
• Audio processed by soundcard
• When soundcard finishes, software distinguishes
  between lower frequency vowels and higher
  frequency consonants
• Compares results with groups of phonemes and then
  to actual words to determine likely match
• Phonemes are the smallest unit of sound that can
  change the meaning of a word, i.e. rat ? bat
• Words are then ordered into most likely sentence
  combinations and output to word processor

15
Downfalls

• Software requires training that can take
  upwards of twelve hours of usage to become fully
  effective
• Human ear and brain have evolved to near
  perfection over millions of years, making it very
  hard to mimic
• Humans are able to discern speech out of more
  than just sound, i.e. out of context, from hand
  and body signals, unconscious lip reading, etc.
• Inconsequential background noise may adversely
  affect softwares ability

16
Man vs. Machine

• TOP Human error vs. machine error in letter
  recognition
• BOTTOM Human error vs. machine error in
  spontaneous speech recognition
• All types of environments tested

17
Two main obstacles

• Software must be able to distinguish between
  background noise and desired input
• Road noise, other people talking, wind, etc
• Must also be able to recognize all different
  types of voices
• High, low, squeaky, raspy, loud, quiet.

18
Biometrics Voice Recognition

• Growing rapidly in corporate and government
  agencies
• Must take completely different approach to
  implementation, software has to be VERY precise
• False accept vs. misprinted word

19
Sound types

• Three types
• Voiced
• Sounds produced by the vibration of the vocal
  cords.
• Unvoiced
• Sounds produced without vibration of vocal cords,
  i.e. sh
• Any sound that does not produce a resonant peak.
• Plosive
• Sounds produced by the release of built up
  pressure in the lungs, i.e. the p in putter.

20
(No Transcript)
21
Approaches to voice authentication

• Template matching
• Speaker-dependent
• Speech analysis
• Speaker-independent

22
Template matching

• User repeats word, series of words, or phrase
  into microphone several times for recording.
• Voice sample is converted from analogue to
  digital for processing.
• Extraction algorithms used to analyze digital
  voice sample looking for unique patterns.
• A mathematical model (stochastic) is created
  which determines certain results from a given
  input according to some statistical distribution.
  
• Hidden Markov model

23
Hidden Markov model

• Variant of a Finite State Machine (FSM)
• An FSM is a model of computation consisting of a
  set of states, a start state, an input alphabet,
  and a transition function that maps input symbols
  and current states to a next state.
• The occurrence of each event in a series of
  random events is dependent on the outcome of the
  adjacent preceding event.
• In any given state, a possible outcome can be
  generated according to its associated probability
  distribution.
• It is only the outcome, not the actual state
  itself, visible to an outside observer, hence
  Hidden Markov model.
• Light switch up or down doesnt matter, only see
  if light is on or off.

24
Template matching (cont.)

• The HMM is used to create statistical profiles of
  each voice sample taken from the user.
• Profiles are compared to look for repeating
  patterns in the users voice that can be used to
  identify his/her voice.
• User is ready to be verified
• Live voice sample compared to user attempting to
  authenticate in database
• Probability score of whether or not user is who
  they claim to be is created
• User accepted or denied based on probability score

25
Feature analysis

• First processes input using Fourier transforms or
  linear predictive coding (LPC).
• Fourier transform is an algorithm used to break
  down signals (sound in the case) into elementary
  components.
• LPC analyzes speech signal by eliminating
  formants (resonances that characterize the vocal
  tract of a person), removing their effects from
  the speech signal, and estimating the intensity
  and frequency of the remaining buzz.
• This process is called inverse filtering and the
  reaming signal is called the residue.

26
Voice authentication obstacles

• Human voice changes with age, sickness, etc.
• Needs to overcome high crossover rate

27
Advantages of Voice Authentication

• Not intrusive
• More natural to speak
• No expensive hardware
• Requires very little memory for voice samples
• Remote Authentication

28
Voice Recognition Uses

• Security
• Financial Transactions
• Account Access
• Funds Transfer
• Bill Payments
• Trading of Financial Instruments
• Credit Card Processing

29
Voice Recognition Uses (contd)

• Authentication
• Senior Citizen Facilities
• Voice recognition systems used in conjunction
  with ID cards are used to let residents into
  their apartments
• Jail System
• Can call up people on parole or on house arrest
  anytime of the day and authenticate if they where
  they should be

30
Voice Recognition Uses (contd)

• Authentication (contd)
• Criminal Cases/Terrorism
• Authentication on videos like the ones where
  Osama Bin Laden appears are authenticated with
  voice recognition to prove that it really is him

31
Uses (cont.)

• Union Pacific Railroad
• New York Town Manor
• Bell Canada
• Password Journal

32
Future Uses

• Small Electronic Devices
• Cell Phones
• Key combination to unlock phones
• PDAs and Laptops
• Same concept could be implemented on these for
  added security

33
Future Uses

• Parking Garage Systems
• Instead of using smart cards that could easily be
  lost, stolen, or even melted, voice recognition
  would take away the possibility of access without
  proper authorization