Strategies for lexical access to very large vocabularies

摘要: Abstract A large vocabulary isolated word recognition system is described on a two pass strategy: hypothesization and verification. Word preselection achieved by segmenting classifying the input signal in terms of 6 broad phonetic classes. To reduce storage computational costs, lexical knowledge organized tree structure where initial common subsequences descriptions are shared, beam-search Dynamic Programming algorithm carries most promising paths only. In second pass, verification, detailed representation phonemic candidates used for estimating likely words. Each candidate modeled graph subword Hidden Markov Models. Again, tree-structure whole subset built online an efficient implementation Viterbi that estimates likelihood candidates. The results show complexity reduction about 73% can be using approach with respect to direct approach, while accuracy remains comparable.