Experiments on the Annotation of Large Quantities of Speech Using the 'ZIP' Dynamic Time-Warping Algorithm

摘要

Research into automatic speech recognition requires large amounts of speech data. This data needs to be carefully annotated (i.e. segmented and labelled) before controlled experiments can be carried out. For example, the researcher may need to know the identity and location of each word in the database. Because of the large amount of data involved it is desirable to have some way of automatically performing this annotation process. Previous experience has demonstrated that a class of algorithms based on dynamic programming can be used to accurately align the timescales of two examples of the same word. An extension of this technique, a dynamic programming algorithm called ZIP, has been shown to be capable of aligning longer examples of speech. This report investigates the feasibility of using the ZIP algorithm as the basis of an automatic annotation system. Experiments are conducted into the possible choice of algorithm parameters and a set of optimal parameters is described. The results obtained demonstrate that this is a viable approach. Methods of automatically determining the success of the algorithm are also considered. (Great Britain)