Different strategies for distribution clustering using discrete, semicontinuous and continuous HMMs in CSR

The authors present an overview of different strategies and refinements to share parameters in HMM models at distribution (state) level for continuous speech recognition, showing the advantages and drawbacks of the different kinds of modeling. They compare them with sharing at the model level, achieving an error reduction close to 20%. Discrete, semicontinuous and continuous HMM models are also compared using these approaches. They consider two ways to smooth discrete distributions (interpolate detailed context dependent with robust context independent) derived from deleted interpolation and co-occurrence smoothing