Tied-Posteriors: A New Hybrid Speech Recognition Technology with Generic Capabilities and High Portability

This paper presents a new method for estimating the emission probabilities of general hybrid connectionist/HMM recognition systems. Contrary to the traditional hybrid approach, where a neu-ral network is used for providing posterior probabilities in order to model the emission probabilities of one-state HMMs, our new tied-posterior approach uses the posterior probabilities resulting from the neural net output in order to replace the Gaussian components of a standard tied-mixture system. This approach allows to use an arbitrary HMM topology with all context-dependency and all clus-tering techniques used in tied-mixture systems. As will be demon-strated in more detail in the paper, this speech recognition architec-ture can be ideally used as generic technology, because it enables the usage of simple straightforward techniques, mainly consisting of training standard neural nets with error-backpropagation and us-ing standard ML techniques for estimating the tied-posterior mix-ture weights. These simple to deploy components lead to a sys-tem yielding very good results even for context-independent mod-els and is superior to the traditional (also easily portable) hybrid posterior approach. Experiments evaluated on the Wall Street Jour-nal (WSJ) database have shown a significant improvement of the recognition rate compared to standard hybrid connectionist/HMM speech recognition systems on this task. 1