We address the problem of automatic phonemic machine labeling of speech database. We are particularly interested in using one hand-labeled utterance of a speaker to automatically label utterances of the same text for a large variety of speakers.
Conventional DTW algorithms used in speech recognition may produce one-to-many or many-to-one alignment between reference and test. Since the difference in speaking rate is globally much lower, the two types of alignment results are not realistic.
We describe a new DTW algorithm which explicitly incorporates in the best path search procedure constraints of phoneme duration. During the alignment of two utterances, the pathes goes across successive phoneme boundaries, both in reference axis and in the test axis. The basic idea is, for each phoneme-to-phoneme transition, weighting the cumulated distance by the duration difference between reference and test phonemes such that its durational deviation from the reference will penalize the path cost. Conventional DTW algorithm is just a special case of the new algorithm.
The results show that: Without phoneme duration constraint, about 11% of machine labeled segments are useless because of abnormal time duration. With phoneme duration constraint this percent drops to 0.05%. At the meantime, due to the improved alignment accuracy we observed phrase recognition error reduction of 5%, on a 33.6 thousand vocabulary task.