HREST performs basic Baum-Welch re-estimation of the parameters of a single HMM using a set of observation sequences. HREST can be used for normal isolated word training in which the observation sequences are realisations of the corresponding vocabulary word.
Alternatively, HREST can be used to generate seed HMMs for phoneme-based recognition. In this latter case, the observation sequences will consist of segments of continuously spoken training material. HREST will cut these out of the training data automatically by simply giving it a segment label.
In both of the above applications, HREST is intended to operate on HMMs with initial parameter values estimated by HINIT.
HREST supports multiple mixture components, multiple streams, parameter tying within a single model, full or diagonal covariance matrices, tied-mixture models and discrete models. The outputs of HREST are often further processed by HEREST.
Like all re-estimation tools, HREST allows a floor to be set on each individual variance by defining a variance floor macro for each data stream (see chapter 8). If any diagonal covariance component falls below 0.00001, then the corresponding mixture weight is set to zero. A warning is issued if the number of mixtures is greater than one, otherwise an error occurs. Applying a variance floor via the -v option or a variance floor macro can be used to prevent this.