One of the most important problems to be faced when building a speech recognizer is the lack of sufficient training data to estimate all of the parameters of the system. One of the most successful techniques addressing this problem is to decrease the number of parameters in the system by the use of classification-tree based acoustic models. In this thesis, several variations on this theme will be presented, the first few ones dealing with trees classifying whole HMMs (tree units), the others with trees classifying states of HMMs (tree states). Results obtained on the Air Travel Information System (ATIS) task show that using classification trees produces systems with up to 90% less models (and proportionally less parameters to estimate) with no loss of performance.
| Identifer | oai:union.ndltd.org:LACETR/oai:collectionscanada.gc.ca:QMM.27540 |
| Date | January 1997 |
| Creators | Lazaridès, Ariane. |
| Contributors | De Mori, Renato (advisor), Normandin, Yves (advisor) |
| Publisher | McGill University |
| Source Sets | Library and Archives Canada ETDs Repository / Centre d'archives des thèses électroniques de Bibliothèque et Archives Canada |
| Language | English |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Format | application/pdf |
| Coverage | Master of Science (School of Computer Science.) |
| Rights | All items in eScholarship@McGill are protected by copyright with all rights reserved unless otherwise indicated. |
| Relation | alephsysno: 001614967, proquestno: MQ37139, Theses scanned by UMI/ProQuest. |
Page generated in 0.0021 seconds