Language Technology : Present and future of automatic speech recognition

1
Language Technology Present and future of
automatic speech recognition

• Jean-Paul Haton
• LORIA-INRIA
• Université Henri Poincaré, Nancy 1
• Institut Universitaire de France
• jph_at_loria.fr

ILSP Oct. 2001 Athens
2
Outline

• Evolution of ASR
• Basic principles
• A major problem robustness
• State of the art
• Future trends and perspectives

3
Applications of ASR

• Data entry voice dictation, etc.
• Access to information travel, banking, portals
• Commands avionics, automobile, etc.
• Speech transcription
• Handicapped
• Speech-to-speech translation

4
Speech recognition technology
5
Principle of Automatic Speech Recognition
6
Bayes Decision Rule
Speech
Parametrization
X x 1T
P(X/W)
Acoustic Models
Search max P(X/W) P(W) W w 1N

Language Models
P(W)
Word sequence
7
Speech modelized by HMMs

• The speech production process is supposed to be
  Markovian

8
Robustness issues in ASR

• Mismatches between training and operating
  conditions can lead to a significant loss in
  performance

9
Main causes of acoustic variation in speech
10
Influence of noise on recognition
11
Approaches to robustness

• Three non-exclusive appoaches
• Signal acquisition and denoising
• Robust parameterization (less sensitive to noise
  variations)
• Model modification adaptation

12
Speech Parametrization

• Use of knowledge from perception,
  psychoacoustics, neurology, etc.
• ear models
• noise masking
• dynamic parameters
• analyses PLP, RASTA, RASTA-PLP, etc.
• Cepstral Normalization

13
Multibands recognition
14
Acoustic Model Adaptation

• Objective
• To derive a new set of models that better match
  the operating environment
• Principle
• To use a small amount of adaptation data
  reflecting the operating environment to modify
  the original models

15
Examples of adaptation methods

• Combination of models PMC (Parallel Model
  Combination)
• Linear transformations of models. Example MLLR
  (Maximum Likelihood Linear Regression)
• MAP (Maximum a posteriori) the a priori
  probability
• of models is not uniform
• Hybridations MLLR/MAP
• Model grouping into classes (eigenvoices, etc.)
• Non-linear transforms (neural networks, etc.)

16
Parallel Model Combination (PMC) Gales, 1995
17
MLLR example
i
u
u
i
a
18
MLLR speaker adaptation
19
State-of-the-art algorithms in speech recognition
Speech input
Acoustic analysis
Phoneme inventory
Pronunciation lexicon
Global search
Recognized Word sequence
20
Future research needs

• Better modelling of speech variability
• (structural and non-structural)
• Better knowledge on speech perception
• and adaptation by humans
• Beyond recognition speech understanding