Technologies for speech applications

1
Chapter 3

• Technologies for speech applications

2
Figure 3.1. Speech Technologies
3
(No Transcript)
4
3.2 Touchtone recognition

• Caller responds to hierarchies of voice menus by
  pressing buttons on the telephone keypad
• Potential problems
• Lost in space
• Time-consuming menus

5
Speech recognition

• Potential problems
• Understandability
• Time-consuming dialogs
• Users may interrupt prompts by barge-in

6
Table 3.1 Speech Reco Engines
 
7
For voice portals

• Continuous speech
• Speaker independent
• Switch vocabularies
• Spontaneous speech
• Multi-threaded

8
Figure 3.4 Speech Recognition
9
Phoneme identification

• Use acoustic model to transform extracted
  features to sequences of phonemes
• Approaches
• Neural networks
• Hidden Markov models

10
Word identification

• Use words in language model to convert sequences
  of phonemes to words
• Two approaches
• Grammars
• N-grams

11
Figure 3.6 Language Model Creation
12
Developers responsibility

• Acoustic model
• Lexicon
• Language model

13
3.4 Voice identification

• General techniques for identifying people
• Something you know
• Something you have
• Something about you

14
Figure 3.2 Speaker registration, identification,
and authentication
15
Voice id technologies are appealing

• Are unobtrusive
• Are location independent
• Require no special equipment
• Replace passwords

16
Why voice technologies fail

• Siblings with similar voice profiles
• Teenage male voice break
• Colds, sore throats, sore lips, etc
• Variety of microphones
• Tape recordings

17
Measuring accuracy of speech id systems
18
3.5 Language identification

• Explicit selectionthe caller speaks the name of
  his or her preferred national language
  ltpromptgt ltspeak xmllang"en-us"gtFor
  service in English, say English lt/speakgt
  ltspeak xmllang"fr"gtpour service en français,
  dites français lt/speakgt lt/promptgt
• Implicit selectiona default language is used
  unless the caller overrides the default with an
  explicit selection ltpromptgt
  ltspeak xmllang"fr"gtBienvenue a portal
  françaislt/speakgt
• ltspeak xmllang"en-us"gtFor
  Service in English, say Englishlt/speakgt
  lt/promptgt 
• Calling area
• Caller  profile
• One number per language.
• Language recognition technology

19
3.6 Word spotting

• Attention word
• User signals that he/she will speak
• Switching contexts
• Suspend current activity and begin a new activity
• Extract critical words
• Security

20
3.7. Language understanding

• Knowledge representation techniques
• Parse trees
• Form templates
• Semantic net
• Approaches for creating knowledge representation
• Parser
• Semantic attachments
• General natural language understanding algorithm

21
3.8 Classification

• Uses
• Classify documents into categories and topics
• Categorize graphical objects
• Replace menu hierarchies in voice applications
• Navigating in large web sites
• Locating a chapter or section of a large document
• Locating a Web site about a specific topic
• Locating descriptions of goods or services in a
  large on-line database
• Example
• How may I help you? from ATT

22
3.9 Dialog management

• Human-driven conversational dialogsthe
  person repeatedly asks a question or speaks a
  command and the computer responds.
• Application-driven conversational
  dialogsthe application repeatedly asks questions
  to solicit answers and instructions from a
  caller.
• Mixed-initiative dialogshuman-driven and
  computer-driven dialogs are combined. The caller
  and computer take turns driving the
  conversations.

23
Figure 3.9 Voice Interpreter and voice Browser
24
3.10 NL processing

• Machine translation
• Word replacement
• Phrase translation
• Full national language translation
• Query generation
• Generate SQL query from knowledge representation
• Summarization
• Generate English summary from knowledge
  representation
• Generation
• Prerecorded sentences
• Templates
• Reversible parse tree

25
Figure 3.7 Speech synthesis
26
Figure 3.8 Concatenative Speech
27
Table 3.2 Concatenative vs parameter-based
speech synthesis
28
3.12 Music synthesis

• Uses
• Branding
• Set the mood
• Signal the caller
• Fundamental part of the dialog
• Approaches
• Prerecord
• Synthesize on the flyMIDI

29
3.13. Tools

• VoiceXML interpreters
• Specification tools
• Call flow tools
• Menu generators
• Form and field generators
• Rehearsal tools
• Logging tools
• Performance measurement summary tools
• System performance tools

30
3.14 Related Technologies

• Distributed Speech recognition
• Noise mitigation
• Noise reduction and cancellation algorithms
• Feature extractionperform on client
• Signal processing algorithms that extract
  essential features from acoustic data
• Multimodal user interfaces
• WML and WAP
• Conversey tags

31
3.15 Key concepts

• Lots of technologies may be useful
• Voice identification
• Speaker registration
• Speaker identification
• Speaker verification
• Speech recognition
• Requires acoustic models, lexicons, language
  models, and grammars
• Speech synthesis
• Synthesis (during development
• Prerecorded (for production)
• Dialog management
• VoiceXML browser