Trajectory behaviour at different phonemic context sizes

Abstract

We propose a piecewise-linear model for the temporal trajectories of Mel Frequency Cepstral Coefficients during phone transitions. As with conventional Hidden Markov Models, the parameters of the model can be estimated for different phonemic context sizes, but our model allows for an intuitive understanding of the impact of context size. We find that the most detailed models, predictably, match the coefficient tracks best - but when data scarcity forces us to use less detailed models, different styles of context modelling (clustered triphones versus biphones) have complementary behaviours. We discuss how this complementarity may be useful for data-efficient ASR.