Experimental Evaluation of Algorithms for Connected Speech Recognition Using Hidden Markov Models

摘要

A method of extracting training utterances from fluent speech and constructing Hidden Markov Models (HMMs) from these templates, known as embedded training, is investigated with a two-level algorithm for connected word recognition. The effects on recognition performance of various HMM training procedures are discussed, and experimental results for native and non-native English speakers are presented. Training on isolated words does not produce models adequate for use in connected word recognition; whether the model was highly unconstrained, to allow for compression of the words in fluent speech, or had a tightly specified transition matrix, to discourage insertion errors, the results are disappointing. The embedded training procedure improves performance, if the Bakis model structure is used; if a full upper-triangular transition probability matrix is used, the performance is far worse than in the isolated-word training case. If the Baum-Welch algorithm is performed after the segmentation procedure, performance improves, but whether this improvement is sufficient to justify the increase in computer time required is questionable.