Americanae. Sistema de Difusión y Recolección de colecciones americanistas

Está en:  › Datos de registro

Maximum entropy-based reinforcement learning using a confidence measure in speech recognition for telephone speech

Identificadores del recurso

http://hdl.handle.net/10533/197895

doi: 10.1109/TASL.2009.2032618

wos: WOS:000278814600013

eissn: 0

issn: 1558-7916

Procedencia

(LA Referencia)

Ficha

Título:: Maximum entropy-based reinforcement learning using a confidence measure in speech recognition for telephone speech
Descripción:: In this paper, a novel confidence-based reinforcement learning (RL) scheme to correct observation log-likelihoods and to address the problem of unsupervised compensation with limited estimation data is proposed. A two-step Viterbi decoding is presented which estimates a correction factor for the observation log-likelihoods that makes the recognized and neighboring HMMs more or less likely by using a confidence score. If regions in the output delivered by the recognizer exhibit low confidence scores, the second Viterbi decoding will tend to focus the search on neighboring models. In contrast, if recognized regions exhibit high confidence scores, the second Viterbi decoding will tend to retain the recognition output obtained at the first step. The proposed RL mechanism is modeled as the linear combination of two metrics or information sources: the acoustic model log-likelihood and the logarithm of a confidence metric. A criterion based on incremental conditional entropy maximization to optimize a linear combination of metrics or information sources online is also presented. The method requires only one utterance, as short as 0.7 s, and can lead to significant reductions in word error rate (WER) between 3% and 18%, depending on the task, training-testing conditions, and method used to optimize the proposed RL scheme. In contrast to ordinary feature compensation and model parameter adaptation methods, the confidence-based RL method takes place in the frame log-likelihood domain. Consequently, as shown in the results presented here, it is complementary to feature compensation and to model adaptation techniques.
Fuente:: IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING; reponame:Artículos CONICYT; instname:CONICYT Chile; instacron:CONICYT
Idioma:: English
Relación:: instname: Conicyt; reponame: Repositorio Digital RI2.0; info:eu-repo/grantAgreement/Fondef/D05I10243; info:eu-repo/semantics/dataset/hdl.handle.net/10533/93477
Ámbito geográfico o temporal:: USA; PISCATAWAY
Autor/Productor:: Molina-Sánchez, Carlos; Huenupan-Quinan, Fernando; Wuth-Sepúlveda, Jorge; Garretón-Vender, Claudio; Becerra-Yoma, Nestor
Editor:: IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
Derechos:: info:eu-repo/semantics/openAccess
Fecha:: 2010
Tipo de recurso:: info:eu-repo/semantics/article; info:eu-repo/semantics/publishedVersion
About:: 2020-01-27T14:04:19Zhttp://www.openarchives.org/OAI/2.0/oai_dc/Artículos CONICYT - CONICYT Chile

oai_dc

Descargar XML

<?xml version="1.0" encoding="UTF-8" ?>

<oai_dc:dc schemaLocation="http://www.openarchives.org/OAI/2.0/oai_dc/ http://www.openarchives.org/OAI/2.0/oai_dc.xsd">
1. <dc:title>Maximum entropy-based reinforcement learning using a confidence measure in speech recognition for telephone speech</dc:title>
2. <dc:creator>Molina-Sánchez, Carlos</dc:creator>
3. <dc:creator>Huenupan-Quinan, Fernando</dc:creator>
4. <dc:creator>Wuth-Sepúlveda, Jorge</dc:creator>
5. <dc:creator>Garretón-Vender, Claudio</dc:creator>
6. <dc:creator>Becerra-Yoma, Nestor</dc:creator>
7. <dc:description>In this paper, a novel confidence-based reinforcement learning (RL) scheme to correct observation log-likelihoods and to address the problem of unsupervised compensation with limited estimation data is proposed. A two-step Viterbi decoding is presented which estimates a correction factor for the observation log-likelihoods that makes the recognized and neighboring HMMs more or less likely by using a confidence score. If regions in the output delivered by the recognizer exhibit low confidence scores, the second Viterbi decoding will tend to focus the search on neighboring models. In contrast, if recognized regions exhibit high confidence scores, the second Viterbi decoding will tend to retain the recognition output obtained at the first step. The proposed RL mechanism is modeled as the linear combination of two metrics or information sources: the acoustic model log-likelihood and the logarithm of a confidence metric. A criterion based on incremental conditional entropy maximization to optimize a linear combination of metrics or information sources online is also presented. The method requires only one utterance, as short as 0.7 s, and can lead to significant reductions in word error rate (WER) between 3% and 18%, depending on the task, training-testing conditions, and method used to optimize the proposed RL scheme. In contrast to ordinary feature compensation and model parameter adaptation methods, the confidence-based RL method takes place in the frame log-likelihood domain. Consequently, as shown in the results presented here, it is complementary to feature compensation and to model adaptation techniques.</dc:description>
8. <dc:date>2010</dc:date>
9. <dc:type>info:eu-repo/semantics/article</dc:type>
10. <dc:type>info:eu-repo/semantics/publishedVersion</dc:type>
11. <dc:identifier>http://hdl.handle.net/10533/197895</dc:identifier>
12. <dc:identifier>doi: 10.1109/TASL.2009.2032618</dc:identifier>
13. <dc:identifier>wos: WOS:000278814600013</dc:identifier>
14. <dc:identifier>eissn: 0</dc:identifier>
15. <dc:identifier>issn: 1558-7916</dc:identifier>
16. <dc:language>eng</dc:language>
17. <dc:relation>instname: Conicyt</dc:relation>
18. <dc:relation>reponame: Repositorio Digital RI2.0</dc:relation>
19. <dc:relation>instname: Conicyt</dc:relation>
20. <dc:relation>reponame: Repositorio Digital RI2.0</dc:relation>
21. <dc:relation>info:eu-repo/grantAgreement/Fondef/D05I10243</dc:relation>
22. <dc:relation>info:eu-repo/semantics/dataset/hdl.handle.net/10533/93477</dc:relation>
23. <dc:rights>info:eu-repo/semantics/openAccess</dc:rights>
24. <dc:coverage>USA</dc:coverage>
25. <dc:coverage>PISCATAWAY</dc:coverage>
26. <dc:publisher>IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC</dc:publisher>
27. <dc:source>IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING</dc:source>
28. <dc:source>reponame:Artículos CONICYT</dc:source>
29. <dc:source>instname:CONICYT Chile</dc:source>
30. <dc:source>instacron:CONICYT</dc:source>
31. <about>
  1. <provenance>
    1. <originDescription altered="" harvestDate="">
      1. <baseURL />
      2. <identifier />
      3. <datestamp>2020-01-27T14:04:19Z</datestamp>
      4. <metadataNamespace>http://www.openarchives.org/OAI/2.0/oai_dc/</metadataNamespace>
      5. <repositoryID />
      6. <repositoryName>Artículos CONICYT - CONICYT Chile</repositoryName>
      </originDescription>
    </provenance>
  </about>
</oai_dc:dc>

xoai

Descargar XML

<?xml version="1.0" encoding="UTF-8" ?>

<metadata schemaLocation="http://www.lyncode.com/xoai http://www.lyncode.com/xsd/xoai.xsd">
1. <element name="dc">
  1. <element name="title">
    1. <element name="none">
      1. <field name="value">Maximum entropy-based reinforcement learning using a confidence measure in speech recognition for telephone speech</field>
      </element>
    </element>
  2. <element name="creator">
    1. <element name="none">
      1. <field name="value">Molina-Sánchez, Carlos</field>
      2. <field name="value">Huenupan-Quinan, Fernando</field>
      3. <field name="value">Wuth-Sepúlveda, Jorge</field>
      4. <field name="value">Garretón-Vender, Claudio</field>
      5. <field name="value">Becerra-Yoma, Nestor</field>
      </element>
    </element>
  3. <element name="description">
    1. <element name="none">
      1. <field name="value">In this paper, a novel confidence-based reinforcement learning (RL) scheme to correct observation log-likelihoods and to address the problem of unsupervised compensation with limited estimation data is proposed. A two-step Viterbi decoding is presented which estimates a correction factor for the observation log-likelihoods that makes the recognized and neighboring HMMs more or less likely by using a confidence score. If regions in the output delivered by the recognizer exhibit low confidence scores, the second Viterbi decoding will tend to focus the search on neighboring models. In contrast, if recognized regions exhibit high confidence scores, the second Viterbi decoding will tend to retain the recognition output obtained at the first step. The proposed RL mechanism is modeled as the linear combination of two metrics or information sources: the acoustic model log-likelihood and the logarithm of a confidence metric. A criterion based on incremental conditional entropy maximization to optimize a linear combination of metrics or information sources online is also presented. The method requires only one utterance, as short as 0.7 s, and can lead to significant reductions in word error rate (WER) between 3% and 18%, depending on the task, training-testing conditions, and method used to optimize the proposed RL scheme. In contrast to ordinary feature compensation and model parameter adaptation methods, the confidence-based RL method takes place in the frame log-likelihood domain. Consequently, as shown in the results presented here, it is complementary to feature compensation and to model adaptation techniques.</field>
      </element>
    </element>
  4. <element name="publisher">
    1. <element name="none">
      1. <field name="value">IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC</field>
      </element>
    </element>
  5. <element name="date">
    1. <element name="none">
      1. <field name="value">2010</field>
      </element>
    </element>
  6. <element name="type">
    1. <element name="none">
      1. <field name="value">info:eu-repo/semantics/article</field>
      2. <field name="value">info:eu-repo/semantics/publishedVersion</field>
      </element>
    </element>
  7. <element name="identifier">
    1. <element name="none">
      1. <field name="value">http://hdl.handle.net/10533/197895</field>
      2. <field name="value">doi: 10.1109/TASL.2009.2032618</field>
      3. <field name="value">wos: WOS:000278814600013</field>
      4. <field name="value">eissn: 0</field>
      5. <field name="value">issn: 1558-7916</field>
      </element>
    </element>
  8. <element name="source">
    1. <element name="none">
      1. <field name="value">IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING</field>
      2. <field name="value">reponame:Artículos CONICYT</field>
      3. <field name="value">instname:CONICYT Chile</field>
      4. <field name="value">instacron:CONICYT</field>
      </element>
    </element>
  9. <element name="relation">
    1. <element name="none">
      1. <field name="value">instname: Conicyt</field>
      2. <field name="value">reponame: Repositorio Digital RI2.0</field>
      3. <field name="value">instname: Conicyt</field>
      4. <field name="value">reponame: Repositorio Digital RI2.0</field>
      5. <field name="value">info:eu-repo/grantAgreement/Fondef/D05I10243</field>
      6. <field name="value">info:eu-repo/semantics/dataset/hdl.handle.net/10533/93477</field>
      </element>
    </element>
  10. <element name="coverage">
    1. <element name="none">
      1. <field name="value">USA</field>
      2. <field name="value">PISCATAWAY</field>
      </element>
    </element>
  11. <element name="rights">
    1. <element name="none">
      1. <field name="value">info:eu-repo/semantics/openAccess</field>
      </element>
    </element>
  12. <element name="language">
    1. <element name="none">
      1. <field name="value">eng</field>
      </element>
    </element>
  </element>
2. <element name="bundles" />
3. <element name="others">
  1. <field name="handle" />
  2. <field name="lastModifyDate">2020-01-27T14:04:19Z</field>
  </element>
</metadata>

Biblioteca AECID
Av. Reyes Católicos, nº 4. 28040 Madrid.
biblio.cooperacion@aecid.es
(+34) 91 583 81 75 - (+34) 91 583 81 64