Finite-State Methods in Automatic Speech Recognition: Difference between revisions
From HLT@INESC-ID
No edit summary |
|||
| Line 2: | Line 2: | ||
== Abstract == | == Abstract == | ||
(to be | This thesis proposes two approaches to address scalability and adaptability problems in weighted finite-state transducer (WFST) approaches to large vocabulary continuous speech recognition. The first one relies on a recognition algorithm which decouples the language model from the other knowledge sources. The second approach is based on a specialized composition algorithm. This algorithm incrementally builds a sequential weighted finite-state transducer representing the composition of the lexicon with the language model, while approximating other optimizations. Being incremental, the algorithm can be embedded in a dynamic speech recognition system. | ||
Both approaches were tested in large vocabulary speech recognition systems. The second one, in particular, was used in a large broadcast news transcription system. A recognition speed improvement of 6 times was observed relative to a previous non-WFST system. | |||
In this thesis various WFST modelling approaches were also pursued. These techniques were applied to two problems in particular: alignment of large speech corpora at both word and phone levels, using phonological rules to model pronunciation variation, and grapheme-to-phone conversion using knowledge-based, data-driven and hybrid approaches. | |||
'''KeywordS:''' Automatic Speech Recognition, Finite-State Methods, Weighted Finite-State Transducers | |||
== Resumo == | == Resumo == | ||
Revision as of 20:37, 3 July 2006
Abstract
This thesis proposes two approaches to address scalability and adaptability problems in weighted finite-state transducer (WFST) approaches to large vocabulary continuous speech recognition. The first one relies on a recognition algorithm which decouples the language model from the other knowledge sources. The second approach is based on a specialized composition algorithm. This algorithm incrementally builds a sequential weighted finite-state transducer representing the composition of the lexicon with the language model, while approximating other optimizations. Being incremental, the algorithm can be embedded in a dynamic speech recognition system.
Both approaches were tested in large vocabulary speech recognition systems. The second one, in particular, was used in a large broadcast news transcription system. A recognition speed improvement of 6 times was observed relative to a previous non-WFST system.
In this thesis various WFST modelling approaches were also pursued. These techniques were applied to two problems in particular: alignment of large speech corpora at both word and phone levels, using phonological rules to model pronunciation variation, and grapheme-to-phone conversion using knowledge-based, data-driven and hybrid approaches.
KeywordS: Automatic Speech Recognition, Finite-State Methods, Weighted Finite-State Transducers
Resumo
(a disponibilizar)
Full Text
- Finite-State Methods in Automatic Speech Recognition (soon)