VoiceDriven Email System

1
Voice-Driven Email System

• Bingchen Li
• Bishad Ghimire
• Puneet Batra

2
Project Discussions

• The discussion will be mainly divided into the
  following parts
• General Introduction
• Hardware Specifications
• Software Specifications

3
General Introduction

• Artificial Intelligence researches
• Speech recognition technologies
• The purpose of our design

4
Design Specifications

• Main purpose is to design and develop
  voice-driven email system
• Its usefulness
• HW, SW and Platform

5
Decomposition

• Systems development more efficient and
  complicated.
• Enables to have a better estimation of the
  current system.
• Is done as
• Client-side Interface
• Speech and Text based HW
• Interface on the Backend

6
The Process

• The SW process is a structured set of activities
  to develop a complicated system
• SW Process Models
• The Rapid Application Development Model (RAD)
• The Fourth Generation Paradigms
• The Concurrent Model
• Project Scheduling

7
Hard Ware
8
HW Specifications

• Foundational Description on TTS
• Our Project is designed to yield an experimental
  but fully operational text/voice system
• Based on a standardized modular interface within
  the design.
• Usefulness

9
Functional Decomposition

• Chipset provides TTS conversion with its
  integrated Text to Speech Synthesizer
• Any English text written to chip is converted
  into speech
• All the data is sent to the chip through serial
  port.

10
Setting the Chipset

• Programming the control panel
• Setting the mode
• Voice type
• Audio Control Registers
• Protocol Options Registers

11
Manufacturers Specifications

• Absolute Maximum Ratings
• Supply voltage, Vcc and AVcc -0.3V to 6.5V
• DC input voltage, Vi -0.3V to Vcc 0.3V
• Operating temp., TA 00C to 700C
•  

12
Contd ..

• DC Characteristics
• TA 0oC to 70 oC
• VCC AVCC AVREF 5v /- 10
• Vss AVss 0v
• XIN 7.35 MHz.

13
Chipset Operation

• Modes
• Text
• Character
• Phoneme
• Real time audio
• Prerecorded audio
• Tone Generator mode
• Idle and standby

14
Exception Dictionary

• Makes possible to alter the way this TTS
  synthesizes
• It can be created and edited for future use
• E.g.
• C(O)N AA
• O after C and before N gets the pronunciation
  AA, the word is cot.

15
How it works

• Some demonstrations
• Text
• Welcome to the Read-Email System Model.
• Phonetic
• w eh l k ah m t uw dh ax r eh d iy m
  ey l s ix s t ax m m aa d ax l

16
Teaching the chip

• Defining the mode
• Text whass
• Phonetic w ae s ah p
• Then,
• Phonetic equi. w ah ts ah p
• Then creat, compile and download.

17
Features of the Chip set

• A full-featured voice synthesizer with unlimited
  vocabulary
• Programmable voice parameters voice, pitch,
  rate, volume and so.
• Non-volatile storage of user settings, dictionary
• DTMF (touch tone) Dialer
• Analog and Digital outputs

18
contd..

• On chip A/D converter
• Very low current drain
• Integrated serial data port.
• Upgradeable and programmable via serial port

19
Systematic flow

• Refresh
• Start with User Inputs
• Authorizing and verifying Entered Information
• Connecting with the Mail Server
• Chip set operation
• Audio Output
• Stops and/or Restart

20
Conclusion

• The chosen chipset has been a good fit
• Successful completion even after programming the
  chips in the phonetic mode
• Working Demo

21
Soft Ware
22
NLP in Information Retrieval (IR)

• What is wrong with todays IR systems ?
• What is Natural Language Processing (NLP) ?
• Different levels of understanding text
• Typical IR system
• Where does NLP come in

23

• Todays Information Retrieval Systems
• google.com, altavista.com, lycos.com
• Mostly Boolean Logic (AND, OR, NOT)
• Some systems use probabilistic and statistic
  logic
• Not very intelligent
• Look at this scenario
• Search Phrase America AND Russia AND Battle
• Search Result A document titled America beats
  Russia in
• battle for gold at summer Olympics

24

• What is Natural Language Processing ?
• Understanding text comes naturally to human
  beings
• We understand text on a number of levels

25

• Different levels of understanding text
• Phonetic
• Morphological
• Syntactic
• Semantic
• Discourse
• Pragmatic

26
A typical IR system
27

• Where does NLP come in ?
• Traditional IR systems use Boolean and
  Probabilistic logic
• In addition, different levels of understanding
  can be
• incorporated
• Number of levels of understanding combined

28
(No Transcript)
29

• Speech Recognition Algorithm
• Hidden Markov model
• Time Sliced Paradigm

30
Hidden Markov model

• What is HMM?
• A triple (?,A,B)
• Usages
• Three problems solved
• The evaluation problem and Forward algorithm
• The decoding problem and Viterbi algorithm
• The learning problem

31
The evaluation problem and Forward algorithm

• What is the evaluation problem?
• How does Forward Algorithm work?
• N
• at1 (j) bj (ot1) S at (i) aij, 1 lt j lt N, 1
  lt t lt T-1
• i1
• What is the benefit?

32
The decoding problem and Viterbi algorithm

• What is the decoding problem?
• How does Viterbi Algorithm work?
• ?t1 (j) bj (ot1) max ?t (i)aij, 1 lt ilt N,
  1 lt t lt T-1 1 lt
  i lt N
• What is the benefit?

33
The learning problem

• What is the learning problem?
• Two main optimization
• Maximum Likelihood (ML)
• Maximum Mutual Information (MMI)

34
Time-Sliced Paradigm

• Introduction
• Objective
• Processes

35
Process Step One

• The spectrum of the input signal is divided into
  slices of equal length along the time-axis, which
  are passed through sequentially through the
  recognition system.

36
Process Step Two
37
Process Step Three
38
Conclusion

• Some final thoughts.