Building Segments with Chunks

Authors: Harald Höge

Abstract:

In HMM-technology state tying is an approved method to achieve reliable estimation of model parameters. This method is based on clustering sub-phonetic units, derived from context dependent phones. Due to the theory of HMMs the clustering algorithm assumes that the feature vectors are statistic independent within a cluster. We derive the sub-phonetic units from tri-phones and call the resulting clusters 'HMM-segments'. In this paper we develop a new clustering algorithm, which is based on the theory of Hidden Chunk Models (HCM). The algorithm takes into account the statistic dependencies of the feature vectors realizing the subphonetic units. We call the resulting segments 'HCM-segments'. Both kinds of segments are modeled with HCMs. With these acoustic models we build two classification systems for context independent phonemes. Using a large Spanish speech database we compare the phoneme error rates achieved with the two kinds of segments. The HCM-segments showed higher performance.


Year: 2014
In session: Spracherkennung
Pages: 16 to 23