Detalhes bibliográficos
| Ano de defesa: |
2005 |
| Autor(a) principal: |
Rezende, Jos? Antonio Moreira de
 |
| Orientador(a): |
Ynoguti, Carlos Alberto
|
| Banca de defesa: |
Meloni, Luis Geraldo Pedroso
,
Fasolo, Sandro Adriano
|
| Tipo de documento: |
Dissertação
|
| Tipo de acesso: |
Acesso aberto |
| Idioma: |
por |
| Instituição de defesa: |
Instituto Nacional de Telecomunica??es
|
| Programa de Pós-Graduação: |
Mestrado em Engenharia de Telecomunica??es
|
| Departamento: |
Instituto Nacional de Telecomunica??es
|
| País: |
Brasil
|
| Palavras-chave em Português: |
|
| Área do conhecimento CNPq: |
|
| Link de acesso: |
http://tede.inatel.br:8080/tede/handle/tede/64
|
Resumo: |
Abstract: The hidden Markov models has become the widely speech recognition framework due to its solid probabilistic approach. Howerver, this robustness is damaged when the Markov chainsare trained under the maximum likelihood methods, which is poor in discriminative power. Until now several efficient techniques and algorithms have been developed to increase the performance of a given speech recognition system. For example, the use of an artificial neural network working with a hidden Markov models, that yields to a Hybrid Model ANN+HMM, wherea neural network is in charge of modeling the acoustic variability and the HMM's are in charge of modeling the temporal variability of a given set of training sentences. Because of the neural network discriminative learning; the is no need to assume for assumption about the statistical distribution of the emission probability. In this dissertation we purpose an investigation about the effects of a manual segmenttation and a simulation of manual segmentations erros on the hybrid ANN+HMM continuos speech recognition system performance, trained under REMAP (recursive estimation and maximization of a posteriori probabilities) algorithm, which improves a posterior probability of the correct model while reducing the probabilities of rival models. |
