Isadora | a Speech Modelling Network Based on Hidden Markov Models

In this paper we present the ISADORA system which provides highly exible speech recognition based on HMM technology together with an hierarchical representation of speech units. Markov model topologies, subword unit inventories, regular grammars expressed in nite-state or phrase structure style, and even the analysis tasks themselves are explicitly represented by the nodes of a large speech unit network. Thus, nothing that can be \said in the language of Markov models" needs to be hard-wired in the program code. In contrast to traditional compiled network recognizers, units, grammars, and tasks may be created or modiied at analysis time, and the outcome of the decoding process is a structured symbolic description of the sensory input. Our architecture has proven extremely useful in prototyping new kinds of subword units. Besides generalized triphones and context-freezing units, a new subword speech unit for automatic speech recognition has been implemented. The so-called polyphones are phone-like units which generalize the well-known concept of triphone units in that more than one left or right context symbol is allowed. Moreover, context items may be of segmental or suprasegmental nature. Moreover, a powerful new training paradigm based on the propagation of statistical parameters through the speech unit network has been introduced. The propagation-based Baum-Welch training algorithm is capable of fast and robust estimation of very large parameter sets | the real-time factor for training is very low (0.3) and independent of utterance duration and model complexity. The paper closes with the presentation of performance gures for numerous continuous speech recognition experiments. Choosing a suitable inventory of polyphones as subword units, a 162-word (or 1081-word, resp.) vocabulary, and using no grammar, speaker-dependent training yielded a word accuracy of 98.3 % (92.4 %). In the speaker-independent mode, accuracies of 91.8 % (84.5 %) have be achieved. This performance is among the best ones reported so far for speaker-independent large-vocabulary continuous speech recognition.