Design of Compact Acoustic Models through Clustering of TiedCovariance Gaussians.pdfVIP

Design of Compact Acoustic Models through Clustering of Tied-Covariance Gaussians Mark Z. Mao and Vincent Vanhoucke Department of Electrical Engineering, Stanford University, CA, USA Nuance Communications, Menlo Park, CA, USA markmao@, vincent@ Abstract also holds for distinct GMMs within a model set: redundancies in the model which are not captured by the phonetic structure We propose a new approach for designing compact acoustic of the language can only be taken advantage of by using a pa- models particularly suited to large systems that combine mul- rameter tying approach which is independent of the phonetic tiple model sets to represent distinct acoustic conditions or lan- structure. guages. We show that Gaussians based on mixtures of inverse In this paper, we will focus on the combination of two shar- covariances (MIC) with shared parameters can be clustered us- ing techniques which operate at the level of the Gaussians and ing an efficient Lloyd algorithm. As a result, more compact the Gaussian parameters, and which make no assumption as to acoustic models can be built by clustering Gaussians across tied which mixture or model set the Gaussians belong to. We will mixtures. In addition, we show that the tied parameters of MIC show that those two techniques work both for sharing Gaussian models can be shared across acoustic models and languages, parameters within a single acoustic model set and for sharing making it possible to build more efficient multi-model systems acro