Monoids: ecient segmental features for speech recognition

Recently, there has been interest in speech recognition with log-linear models that use features for whole segments, for example, words. e segmentation is oŸen taken from a conventional speech recogniser. However, this limits the performance ofmoving to a newmodel. An alternative is to nd the optimal segmentation. is requires acoustic features for all possible segments, which a recently proposed method extracts e›ciently. It shares computation between features for segments with the same start time. is is useful when all segments are considered, but feature extraction still takes quadratic time in the length of the utterance. A more realistic strategy for decoding would prune the hypothesis space. is report therefore proposes a new, more žexible class of features. When features for all segments are required, extracting them has the same time complexity, but when only a limited number of segments are considered, they allow more re-use of computation. A specic subclass of features of interest derives from the total weight of a hidden Markov model (hmm) or a similar nite-state model. is report shows how to compute scores e›ciently for such a nite-state model with weights in any semiring.