Development and application of an accurate and flexible automatic aligner

Abstract

Abstract. Large databases are useful tools for speech technology research. Their usefulness is greatly enhanced if the data is annotated with time aligned labels. This is expensive and time consuming and has lead to the investigation and development of automatic aligners. This paper reports on an automatic aligner developed initially to solve the problem of annotating a large database within a set period of time. While developing the aligner, we investigated the importance of the models, the use of manual labels to bootstrap the system, and the role of the dictionary in the effectiveness of the aligner, and found that each had a contribution to make. The aligner produced was tested on unseen data to gauge its accuracy before being applied as a tool to annotation of a large amount of data. The aligner was developed in a way that facilitates its use in other applications.

Precision of phoneme boundaries derived using hidden Markov models

“…The most common method of assessing labelling accuracy is to compare the automatic system's labels with a 'goldstandard' set defined by one or more phoneticians [1,2], but this is inherently subjective and requires much manual effort. To avoid these difficulties, we have developed an objective measure of the quality of the alignment system, that operates without reference to human interpretations.…”