The SPEERAL Decoder

1
The SPEERAL Decoder

• NOCERA Pascal
• Laboratoire d Informatique d Avignon
• AGROPARC
• BP 1228, 84911 AVIGNON Cedex 9
• Tel [phone]
• E-mail pascal.nocera_at_lia.univ-avignon.fr

2
The SPEERAL System
Stochastic approach Find the best
hypothesis among all the possible hypotheses with
the A algorithm.
3
The SPEERAL System
Stochastic approach
4
Acoustic Models
Hidden Markov Models Gaussian Mixture
Models Contextual Models (Phonemes)
5
Acoustic Model Toolkit

• Parameterization program
• Text to phone program
• Alignment program
• HMM learning program
• Supervised and unsupervised Model Adaptation
• MLLR
• MAP
• Structural Model Space Transformation

6
Linguistic Models

• Stochastic Language Models
• N-grams
• Class based language models

7
Linguistic Model Toolkit

• Text Normalization Tools
• Language Model Training
• CMU toolkit
• SRI toolkit
• ATT toolkit
• Language Model Compilation
• Lexicon Compilation

8
Standard A algorithm

•  best-first  search algorithm
• Extend the best path to generate new candidates
• Assign a score F(x) to all explored path
• g(x) combines Language Model and acoustic scores
• h(x) estimates the probability of the best
  extension
• Keep the list of explored paths as a priority
  queue
• When the best path reaches end then stop

9
Standard A algorithm (2/2)

• Requires an admissible heuristic function
• h(x) underestimates the true remaining cost path
  (the more accurate the better).
• Heuristics samples
• h(x) 0
• Breadth-First search
• h(x) true remaining cost (i.e. F(x) never
  changes)
• Deterministic search

10
The SPEERAL System

• Language model
• Stochastic n-gram LM (n3)
• Lexical, phonetic and acoustic knowledge source
• Acoustic model (HMM, )
• Decoding vocabulary (lexicon)
• Input signal ? Phoneme lattice
• ( p, beg, end, sc ) with score sc P(X
  beg..end/p)
• /

11
Sounding function h

• Remaining path estimation
• Acoustic score only
• Computed with a backward Viterbi, during the
  phoneme lattice generation
• Heuristic admissibility
• Underestimate remaining cost no LM information
• Cannot be true cost (lack of LM information)

12
Lexicon

• Prefix-tree organization
• Widely applied
• Compact representation
• search effort occurs at word begin

13
Search space

• Phoneme lattice
• Concatenation of lexical trees

14
LM look-ahead

• Word anticipation
• n is a lexicon node
• wn is any leaf (i.e. word) of the sub-tree
  starting at n
• P(n/...wi-2 w i-1) Part_LM(n, wi-2 wi-1 )
• Part_LM(n, wi-2 w i-1 ) maxWnP(wn/wi-2 wi-1)
• ? Paths leading to improbable words are early
  penalized

15
Start-synchronous tree

• Asynchronous search
• The search processes the same part (lexicon) with
  a different history.
• With start-synchronous capabilities
• Most advanced path can be reused when encountered
  twice.
• For each frame x, the lexicon starting at x is
  stored.
• Only the deepest nodes (or leaves) are stored.

16
Principle (1/5)
Frame t
Frame 0
Deepest lexicon nodes at frame 0
Deepest lexicon nodes at frame t
p
W
p
1
2
3
p
1
p
1
p
W
p
p
3
3
2
1
p
p
2
1
17
Principle (2/5)
Frame t
Frame 0
p
2
p
W
p
1
2
3
p
1
p
3
p
1
p
W
p
p
3
3
2
1
p
p
2
1
18
Principle (3/5)  
Frame t
Frame 0
W
2
p
2
p
W
p
1
2
3
p
1
p
3
p
1
p
W
p
p
3
3
2
1
p
p
2
1
19
Principle (4/5) 
Frame t
Frame 0
W
2
p
2
p
W
p
1
2
3
p
1
p
3
p
1
p
W
p
p
3
3
2
1
p
p
2
1
20
Principle (5/5) . 
Frame t
Frame 0
p
Frame tn
1
p
2
W
2
p
2
p
1
p
W
p
1
2
3
p
1
p
p
3
2
p
1
p
W
p
p
3
3
2
1
p
p
2
1
21
Search space pruning

• Optimization
• If two candidates end with the same 3 words, only
  the best is kept.
• Cut
• Short candidates are dropped when their distance
  increase too much with the deepest.

22

• ASR Output
• 1 best hypothesis
• N best hypothesis
• word graph

• Applications
• Transcription
• Question answering
• Named entities extraction
• Information Retrieval
• Call-type classification

23
French Broadcast News Campain ESTER
Broadcast News (1h long show)
24
System Description

• Acoustic Models
• 10k HMM contextual
• 3.6k states
• 230k gaussian
• Lexicon 65K Words
• Language model Combination
• (Le Monde 87-02, 0.41)
• (Le Monde 02-03, 0.24)
• (ESTER, 0.35)

25
Results and Demonstration

• WER 25 (10 RT)
• Demonstration on TV