Basic concept for how for recognize the speech. Using for Artificial Intelligence

1. SPEECH RECOGNITION: ADVANCED
TOPICS
True, their voice-print machine was unfortunately a crude
one. It could discriminate among only a few frequencies, and
it indicated amplitude by indecipherable blots. But it had
never been intended for such vitally important work.
--- Aleksandr I. Solzhenitsyn, The First Circle, p. 505

2. INTRODUCTION
• keju civil service examinations of Imperial China lasted almost 1300 years,
• from the year 606 until it was abolished in 1905. this exam is taken for finding
the high rank officer for china.
• This keju examination process are called as keju algorithm, which is
incremental and progressive process of finding the individual candidate.
• Keju algorithm are also used for speech recognition. Which is very expensive
algorithm for speech recognition.
• So Author introduce the `Multipass Decoding Algorithm` which is efficient but
dumper decoding algorithm to produce shortlists of potential candidates

3. TOPIC OF THIS CHAPTER
This chapter author introduce several methodology to discuss
1. Multi pass decoding algorithm
Dumper decoding algorithm produce the shortlist of probabilistic path, in
which finding and rescore the potential candidate.
2. Context dependent acoustic model
Smart process for create the large vocabulary of speech recognition.
3. Discriminative training
4. Modeling the variation

4. MULTI PASS DECODING
For multi pass decoding, we have use stack decoder or A* decoder, which divide the
multi pass decoding in to two stage.
First Stage we have use efficient knowledge sources to perform optimal search and
second stage using sophisticated and slower decoding algorithm to reduce the search
space.
1. N-BEST LISTS
2. WORD Lattices

5. N-BEST LISTS
• This algorithm is modification of `Viterbi algorithm` to return the N-Best Sentence
from given speech.

6. LIMITATION OF N-BEST LIST
1. One problem with an N-best list is that when N is large, listing all the sentences is
extremely inefficient.
2. Another problem is that N-best lists don’t give quite as much information as we
might want for a second-pass decoder.

7. WORD LATTICE
The output of the first pass decoder is usually a more sophisticated representation called a
word lattice(Murveit et al., 1993; Aubert and Ney, 1995). A word Lattice is a directed graph
that efficiently represents much more information about possible word sequence. There is
two part of directed graph,
• Node
• Arcs
Nodes in the graph are word
Arcs are transitions between word
Arch Represent word hypotheses and nodes are points in time

8. WORD LATTICES

9. N-BEST LIST AND WORD LATTICES GOAL
N-Best List and Word Lattices goal is to
rescoring the probabilistic candidate and
replace with 1-best utterance with a different
utterance.

10. A∗ (‘STACK’) DECODING
The A∗ decoding algorithm allows us to use the complete forward probability, avoiding the
Viterbi approximation
A∗ decoding also allows us to use any arbitrary language model.
Thus A∗ is a one-pass alternative to multi-pass decoding
The A* decoding algorithm is a best first search of the tree that implicitly defines the
sequence of allowable word in the language. This algorithm has tow parts,
1. The root: start node on the left or START point of the path
2. Leaf: Difference path of the probabilistic candidate, each Leaf define the on sentence ot
the language and create the path

11. A∗ (‘STACK’) DECODING

12. A∗ (‘STACK’) DECODING
Priority Queue
The A* decoder must thus find the path (word sequence), the root to a leaf which has
the highest probabilities, where the path probability is defined as the product of its
language model probability(prior) and its acoustic match to the data.
Fast match
A Fast match is used to select the likely next words. A fast match is one of a class of
heuristics designed to efficiently winnow down the number of possible following words,
often by computing some approximation to the forward probability.

13. A∗ (‘STACK’) DECODING

14. STACK DECODING

15. END FOR TODAY……………………..