Objective Optimisation of Automatic Speech-to-Phoneme Alignment Systems

This paper presents techniques for objective characterisation of Automatic Speech-to-Phoneme Alignment (ASPA) systems, without the need for human-generated labels to act as a benchmark. As well as being immune to the effects of human variability, these techniques yield diagnostic information which can be helpful in the development of new alignment systems, ensuring that the resulting labels are as consistent as possible. To illustrate this, a total of 48 ASPA systems are used, including three front-end processors. For each processor, the number of states in each phoneme model, and of Gaussian distributions in each state mixture, are adjusted to generate a broad variety of systems. The results are compared using a statistical measure and a model-based Bayesian Monte-Carlo approach. The most consistent alignment system is identified, and is (as expected) in close agreement with typical “baseline” systems used in ASR research.