{"id":{"repo_id":"aachen","oai_identifier":"oai:publications.rwth-aachen.de:60519"},"canonical_url":"https://search.dev.ndltd.org/etd/aachen/oai:publications.rwth-aachen.de:60519","repository":{"repo_id":"aachen","name":"RWTH Aachen University","base_url":"https://publications.rwth-aachen.de/oai2d"},"display":{"title":"Diskriminative Modellkombination in Spracherkennungssystemen mit großem Wortschatz","abstract":"In this work, the theory of Discriminative Model Combination, DMC, is developed and implemented for large vocabulary continuous speech recognition. DMC is based on a discriminative training of the free parameters of distributions belonging to the exponential family. It is independent of the combined models and allows for the automatic combination of any set of models of any kind. The smoothed empirical word error rate is exploitet as optimization criterion. Using the DMC method the LVCSR system of Philips Research Laboratories Aachen could be improved significantly on the Wallstreet-Journal Task and on the Broadcast-News Task. It is shown experimentally, that the log-linear functional form of the model combination outperforms a linear form and a simple voting scheme. In addition, the smoothed empirical word error rate criterion is reformulated in a way, which allows to compute the weights of the model combination in a closed form. DMC is independent of the hierarchical level of the classification task and of the applied models. This is shown by deriving algorithms for the balancing of the language model weight, for the log-linear combination of any acoustic and language models, for the log-linear combination of multilingual phoneme models and phoneme class models, for the balancing of the influence of transition probabilities and the emission distribution of the Hidden-Markov-Models, for the estimation of the free parameters of multivariate gaussian distributions as well as for the log-linear language model interpolation.","abstract_html":"In this work, the theory of Discriminative Model Combination, DMC, is developed and implemented for large vocabulary continuous speech recognition. DMC is based on a discriminative training of the free parameters of distributions belonging to the exponential family. It is independent of the combined models and allows for the automatic combination of any set of models of any kind. The smoothed empirical word error rate is exploitet as optimization criterion. Using the DMC method the LVCSR system of Philips Research Laboratories Aachen could be improved significantly on the Wallstreet-Journal Task and on the Broadcast-News Task. It is shown experimentally, that the log-linear functional form of the model combination outperforms a linear form and a simple voting scheme. In addition, the smoothed empirical word error rate criterion is reformulated in a way, which allows to compute the weights of the model combination in a closed form. DMC is independent of the hierarchical level of the classification task and of the applied models. This is shown by deriving algorithms for the balancing of the language model weight, for the log-linear combination of any acoustic and language models, for the log-linear combination of multilingual phoneme models and phoneme class models, for the balancing of the influence of transition probabilities and the emission distribution of the Hidden-Markov-Models, for the estimation of the free parameters of multivariate gaussian distributions as well as for the log-linear language model interpolation.","abstract_has_math":false,"creators":["Beyerlein, Peter"],"institution":"Publikationsserver der RWTH Aachen University","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":["Ney, Hermann"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2001,"date_issued":"2001","date_published":"2001","updated_at":"2026-07-30T19:42:56Z","subjects":["info:eu-repo/classification/ddc/620","Ingenieurwissenschaften","Automatische Spracherkennung","Wortschatz","Übertragungsfehler","Bitrate","Diskriminanzanalyse","Maximum-Entropie-Methode"],"languages":["ger"],"rights":["info:eu-repo/semantics/openAccess"],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-122224%22"],"render_values":[{"text":"https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-122224%22","href":"https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-122224%22","code":true}]}]},"links":{"outbound_url":"https://publications.rwth-aachen.de/record/60519","outbound_label":"Repository record","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Ney, Hermann"]},{"key":"dc:creator","label":"Author","values":["Beyerlein, Peter"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:coverage","label":"Dc Coverage","values":["DE"]},{"key":"dc:date","label":"Dc Date","values":["2001"]},{"key":"dc:publisher","label":"Institution","values":["Publikationsserver der RWTH Aachen University"]},{"key":"dc:relation","label":"Dc Relation","values":["info:eu-repo/semantics/altIdentifier/urn/urn:nbn:de:hbz:82-opus-1254"]},{"key":"dc:type","label":"Dc Type","values":["info:eu-repo/semantics/doctoralThesis","info:eu-repo/semantics/publishedVersion"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["info:eu-repo/classification/ddc/620","Ingenieurwissenschaften","Automatische Spracherkennung","Wortschatz","Übertragungsfehler","Bitrate","Diskriminanzanalyse","Maximum-Entropie-Methode"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["ger"]},{"key":"dc:rights","label":"Dc Rights","values":["info:eu-repo/semantics/openAccess"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://publications.rwth-aachen.de/record/60519","https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-122224%22"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["In this work, the theory of Discriminative Model Combination, DMC, is developed and implemented for large vocabulary continuous speech recognition. DMC is based on a discriminative training of the free parameters of distributions belonging to the exponential family. It is independent of the combined models and allows for the automatic combination of any set of models of any kind. The smoothed empirical word error rate is exploitet as optimization criterion. Using the DMC method the LVCSR system of Philips Research Laboratories Aachen could be improved significantly on the Wallstreet-Journal Task and on the Broadcast-News Task. It is shown experimentally, that the log-linear functional form of the model combination outperforms a linear form and a simple voting scheme. In addition, the smoothed empirical word error rate criterion is reformulated in a way, which allows to compute the weights of the model combination in a closed form. DMC is independent of the hierarchical level of the classification task and of the applied models. This is shown by deriving algorithms for the balancing of the language model weight, for the log-linear combination of any acoustic and language models, for the log-linear combination of multilingual phoneme models and phoneme class models, for the balancing of the influence of transition probabilities and the emission distribution of the Hidden-Markov-Models, for the estimation of the free parameters of multivariate gaussian distributions as well as for the log-linear language model interpolation."]},{"key":"dc:source","label":"Dc Source","values":["Aachen : Publikationsserver der RWTH Aachen University 125 S. : graph. Darst. (2001). = Aachen, Techn. Hochsch., Diss., 2000"]},{"key":"dc:title","label":"Title","values":["Diskriminative Modellkombination in Spracherkennungssystemen mit großem Wortschatz"]}]}],"canonical_facts":{"dc:contributor":["Ney, Hermann"],"dc:coverage":["DE"],"dc:creator":["Beyerlein, Peter"],"dc:date":["2001"],"dc:description":["In this work, the theory of Discriminative Model Combination, DMC, is developed and implemented for large vocabulary continuous speech recognition. DMC is based on a discriminative training of the free parameters of distributions belonging to the exponential family. It is independent of the combined models and allows for the automatic combination of any set of models of any kind. The smoothed empirical word error rate is exploitet as optimization criterion. Using the DMC method the LVCSR system of Philips Research Laboratories Aachen could be improved significantly on the Wallstreet-Journal Task and on the Broadcast-News Task. It is shown experimentally, that the log-linear functional form of the model combination outperforms a linear form and a simple voting scheme. In addition, the smoothed empirical word error rate criterion is reformulated in a way, which allows to compute the weights of the model combination in a closed form. DMC is independent of the hierarchical level of the classification task and of the applied models. This is shown by deriving algorithms for the balancing of the language model weight, for the log-linear combination of any acoustic and language models, for the log-linear combination of multilingual phoneme models and phoneme class models, for the balancing of the influence of transition probabilities and the emission distribution of the Hidden-Markov-Models, for the estimation of the free parameters of multivariate gaussian distributions as well as for the log-linear language model interpolation."],"dc:identifier":["https://publications.rwth-aachen.de/record/60519","https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-122224%22"],"dc:language":["ger"],"dc:publisher":["Publikationsserver der RWTH Aachen University"],"dc:relation":["info:eu-repo/semantics/altIdentifier/urn/urn:nbn:de:hbz:82-opus-1254"],"dc:rights":["info:eu-repo/semantics/openAccess"],"dc:source":["Aachen : Publikationsserver der RWTH Aachen University 125 S. : graph. Darst. (2001). = Aachen, Techn. Hochsch., Diss., 2000"],"dc:subject":["info:eu-repo/classification/ddc/620","Ingenieurwissenschaften","Automatische Spracherkennung","Wortschatz","Übertragungsfehler","Bitrate","Diskriminanzanalyse","Maximum-Entropie-Methode"],"dc:title":["Diskriminative Modellkombination in Spracherkennungssystemen mit großem Wortschatz"],"dc:type":["info:eu-repo/semantics/doctoralThesis","info:eu-repo/semantics/publishedVersion"]},"updated_at":"2026-07-30T19:42:56Z"}