TY - JOUR AU - Deng, L. PY - 1991/09/01 Y2 - 2024/03/28 TI - Hierarchical nonstationarity in a class of doubly stochastic models with application to automatic speech recognition JF - Canadian Acoustics JA - Canadian Acoustics VL - 19 IS - 4 SE - Proceedings of the Acoustics Week in Canada DO - UR - https://jcaa.caa-aca.ca/index.php/jcaa/article/view/695 SP - 113-114 AB - Introduces the concept of two-level (global and local) hierarchical nonstationarity for describing the complex, elastic, and highly dynamic nature of speech signals. A general class of doubly stochastic process models are developed to implement this concept. In this class of models, the global nonstationarity is embodied through an underlying Markov chain (or any other scheme capable of providing nonlinear time warping mechanisms) which governs the evolution of the parameters in a set of output stochastic processes. The local nonstationarity is realized by assuming state-conditioned, time-varying first and second order statistics in the output data-generation process models. To provide practical algorithms for speech recognition which allow the model parameters to be reliably estimated, the local nonstationarity is represented in a parametric form. Simulation results demonstrated close fitting of the model to the actual speech data. Results from speech recognition experiments provided evidence for the effectiveness of the model in comparison with the standard HMM, which is a degenerated case-with single-level nonstationarity-of the proposed model ER -