Spoken Language Dialog Systems

1
Spoken Language Dialog Systems

• Robert DaleRobert.Dale_at_mq.edu.au

2
The Architecture of an SLDS
Speech Recognition
Speech Synthesis
Language Understanding
Language Generation
Dialog Management
Database
3
Typical Systems Today

• Speech Recognition accuracy of 85-90
• Language Processing Very limited
• Dialog Management strongly-managed dialog flow
• Language Generation Very trivial
• Speech Synthesis high quality for short
  utterances

4
Speech Recognition

• Commercially-deployed systems use grammar-based
  recognition that closely constrains what can be
  said
• Laboratory systems use n-gram language models
  that can recognise a much broader range of
  language

5
Language Understanding

• Effectively non-existent in commercially deployed
  systems recognition of predefined word
  sequences leads directly to a corresponding
  action
• Laboratory systems use syntactic analysis of
  recognized word sequence followed by some kind of
  semantic analysis coverage always much more
  limited than the sequences of words that are
  recognized

6
Dialog Management

• Commercially-deployed systems use finite-state
  dialog models that closely prescribe the flow of
  the dialog
• Laboratory systems use more complex models that
  allow some reasoning about the state of the
  dialog to provide dynamic flexibility

7
DA Call Flow
Welcome to Telstras Directory Assistance
System.
1
Retry Appropriate?
3
2
No Rec
Sorry, I didnt hear you. Please wait until the
prompt is finished before speaking.
What name, please?
4
Name recognised
What location, please?
5
Agent
6
7
Did you want ?
No Rec
Sorry, I didnt hear you. Please wait until the
prompt is finished before speaking.
Retry Appropriate?
13
No
No
The number is
Yes
8
No
2nd Quneeded?
High Charge?
No Main Number
9
Yes
Yes
Yes
14
For attemptedconnection
Rec Ok
10
I heard Is that right?
Yes
12
11
MainNumber?
Please choose or or another .
No Rec
Yes or 1
No
We dont have Shall I give you ...?
15
No
16
Answered?
No
To Agent
18
IBM
Yes
17
Main Number Exists
Connect
8
Language Generation

• Commercially deployed systems use predetermined
  prompts at each dialog state
• Laboratory systems reason about what to say and
  how to say it in order to construct an
  appropriate utterance

9
Speech Synthesis

• Commercially deployed systems generally use
  recorded voice to ensure quality, with TTS being
  used for situations where recording is too
  expensive or difficult
• Laboratory systems use unit concatenation
  techniques that provide almost human-like quality
  in some cases, but with little opportunity for
  prosodic control

10
Where Does This Leave Us?

• Commercially deployed systems are best thought of
  as smart answering machines
• Laboratory systems attempt to emulate human
  conversational behaviour but risk raising
  expectations too high
• The solution lies in the middle
  semi-conversational systems?