Voice as a User Interface Tuning Speech

1
Voice as a User InterfaceTuning Speech

• Baiju D. Mandalia, PhD
• Senior Technical Staff Member, IBM Corporation

2
Voice as a User Interface

• Human factors standards
• Error handling
• Prompt Design
• Dialog Design
• Grammar Design
• Tuning voice applications

3
ETSI ES 202 077 for spokencommands

• Context-independent common commands (like Main
  Menu, Operator, Goodbye etc.)
• Context dependent common commands (Help, Repeat)
• Core commands (Yes, No, Stop)
• Digits (including zero, Oh, double etc)
• Name and Digit dialing (like Home, Work, Mobile)
• Basic Call Handling (Answer, Divert all calls
  to, Transfer etc.)
• Media control (Play, Pause, Continue etc.)
• Browseable List Navigation (Next, Continue,
  Details)
• Editing Commands (Delete, Save, Record etc.)
• Device settings (Volume up/ Louder, Volume
  down/Quieter)
• Word-spotting mode (Wake-up- to activate other
  modal functions in above list)

4
Dialog Components for handling errors
5
Prompt Design

• Pre-recorded vs Text to Speech
• Guiding the caller
• SSML
• Dictionaries

6
Dialog Design

• Consistency
• Make sure caller experience is uniform during
  call
• Barge-in
• Define dialog based on use of better interaction
  and barge-in by experienced users
• Using Nbest
• Exploit nbest to provide disambiguation with
  complex tasks

7
Grammar Design

• Unknown pronunciations
• Acoustic Confusability
• Grammar Coverage
• Complexity
• Dynamic Application Development
• Weighting of more common words

8
Tuning Voice Applications

• Timeouts
• Lexicons
• Weighting
• Confidence levels
• Speed vs Accuracy
• Sensitivity
• Acoustic model adaptation

9
WebSphere Voice Server Tuning Tools
10
Tuning tools features

• Eclipse based
• User friendly , graphical interface
• Tightly integrated for repetitive testing
• Assistive tools like pronunciation builder for
  tuning

11
Validating grammars on the MRCP Server
12
Enumerating a grammar (random)
13
Testing grammars with text
14
Pronunciation Builder
15
Testing grammars with speech
16
Voice Trace Analyzer for Tuning

• Set the Voice Server trace specification.
• Run your voice application (generate trace data).
• Run the WVS Collector tool.
• (optional for Integrated Runtime Environment)
• Import the data into the Voice Trace Analyzer.

17
Voice Trace Analyzer Views
18
Transcriptions within tool

• In Grammar
• Accuracy
• Correct Accept (CA)
• Correct Reject (CR)
• False Accept (FA)
• FA-In
• FA-Out
• False Rejects (FR)

19
References

• WebSphere Voice Server
• http//www.ibm.com/software/pervasive/voice_serve
  r
• WebSphere Voice Server Information Center
• http//publib.boulder.ibm.com/infocenter/pvcvoice
  /51x/index.jsp
• WebSphere Voice Zone
• http//www.ibm.com/developerworks/websphere/zones
  /voice
• IBM WVS for Multiplatforms V5.1.1/V5.1.2 Handbook
• http//www.redbooks.ibm.com/abstracts/sg246447.htm
  l?Open
• Speech User Interface Guide
• http//www.redbooks.ibm.com/redpieces/abstracts/r
  edp4106.html?Open

20
(No Transcript)