8
Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO, and ISIP_LM_TESTER Progress on the Speech Recognition Search Demo

Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Embed Size (px)

Citation preview

Page 1: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Jun-Won Suh Intelligent Electronic Systems

Human and Systems EngineeringDepartment of Electrical and Computer Engineering

ISIP_VERIFY, ISIP_DECODER_DEMO, and ISIP_LM_TESTER

Progress on the Speech Recognition Search

Demo

Page 2: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Page 2 of 7Research Progress: Jun-Won Suh

Overview

• Search Decoding Demo

• Language Model Tester

• Verify Utility

• Thesis Topic

• Speaker Verification using HMM, SVM, RVM

Page 3: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Page 3 of 7Research Progress: Jun-Won Suh

Search Decoding Demo

• Previous Problem

Possible word sequences disappear

Overlapping over Hypothesis and possible word sequences

N-gram was not working

Need to add the context dependent and independent mode

• Current Problem

• Lattice rescoring gives an segmental fault

Page 4: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Page 4 of 7Research Progress: Jun-Won Suh

Language Model Tester

• Current Language Model Tester

Using the IHD grammar, generate the symbols or graphs depending on searchlevel object @ Sof v1.0 @

@ AnnotationGraph 0 @

id = "hello_world";

type = "ORTHOGRAPHIC";

anchors = {

{0}, {id = "hello_world:Anchor1";

offset = 0;

unit = "seconds";

anchored = false;

}

}, {

{1}, {id = "hello_world:Anchor2";

offset = 0;

unit = "seconds";

anchored = false;

}

annotations = {

{0}, {1}, {id = "hello_world:Annotation1";

type = "!SENT_DELIM !DUMMY ONE ";

channel_index = 0;

features = {

load_factor = 0.75;

table = {

{"level"}, {"word"} }

}

Page 5: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Page 5 of 7Research Progress: Jun-Won Suh

Language Model Tester

• Add Functionality: Parsing

To take an existing sequence of symbols and determine if it was a permissible by the given grammar

- Language Model

- Hierarchical Digraph

- SearchLevel

- Vector<SearchSymbol>

- Vector<Digraph<Ulong>>

- Digraph<Ulong>

- Vertices

- Arcs

@ Sof v1.0 @

@ LanguageModel 0 @

algorithm = "IHD";

implementation = "IHD";

h_digraph = {

level_index = 0;

search_tag = "word";

search_symbols = {

"!DUMMY"

}, {

"!SENT_DELIM"

}, {

"EIGHT"

}, {

"FIVE"

}, {

"FOUR"

}, {

"NINE"

}, {

"OH"

}, {

"ONE"

}, {

"SEVEN"

}, {

"SILENCE"

}, {

"SIX"

}, {

"THREE"

}, {

"TWO"

}, {

"ZERO"

};

search_models = {

weighted = true;

vertices =

{0, {1}},

{1, {0}},

{2, {7}},

{3, {0}},

{4, {9}},

{5, {1}},

{6, {12}},

{7, {11}},

{8, {4}},

{9, {3}},

{10, {10}},

{11, {8}},

{12, {2}},

{13, {5}},

{14, {6}},

{15, {13}};

arcs =

{S, 0, 0},

{0, 1, 0},

{1, 6, 0},

{1, 7, 0},

{1, 8, 0},

{1, 9, 0},

Each word sequences tokenized

Find corresponding Search Symbol index in Vertex

Each word sequences will be stored in pair object

Compare with Arc

Write the result in Sof file

Page 6: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Page 6 of 7Research Progress: Jun-Won Suh

Language Model Tester

• Plans for parsing each Search Level

s t

one

two

st

w oh

t ow

n

dummy dummy

dummydummy

s s1 t

s t

s2 s3

s1 s3 s4

2 search level

1 search level

0 search level

Page 7: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Page 7 of 7Research Progress: Jun-Won Suh

Verify Utility: Big Possible Thesis Topic!!

• Had many discussion to combine the HMM, SVM, and RVM

• Set up the parameters

• Need More Specific Thesis Topic

• Need to have results until End of June

• Change the SVM and RVM parameters to improve ERR??

Current Activity for Thesis

- Read the “Speaker Recognition: A Tutorial”, J. P. Campbell

- Read the “Introduction to Support Vector Machines”, P.S. Sastry

Page 8: Jun-Won Suh Intelligent Electronic Systems Human and Systems Engineering Department of Electrical and Computer Engineering ISIP_VERIFY, ISIP_DECODER_DEMO,

Page 8 of 7Research Progress: Jun-Won Suh

Need More Research

• Getting used to familiar with IFC and utility

• Need more language model knowledge