Autoregressive HMM speech synthesis

Autoregressive HMM modeling of spectral features has been proposed as a replacement for standard HMM speech synthesis. The merits of the approach are explored, and methods for enforcing stability of the estimated predictor coefficients are presented. It appears that rather than directly estimating autoregressive HMM parameters, greater synthesis accuracy is obtained by estimating the autoregressive HMM parameters by using a more traditional HMM recognition system to compute state-level posterior probabilities that are then used to accumulate statistics to estimate predictor coefficients. The result is a simplified mathematical framework that requires no modeling of derivatives and still provides smooth synthesis without unnatural spectral discontinuities. The resulting synthesis algorithm involves no matrix solves and may be formulated causally, and appears to result in quality very similar to that of more traditional HMM synthesis approaches. This paper describes the implementation of a complete Autoregressive HMM LVCSR system and its application for synthesis, and describes the preliminary synthesis results.